R
e
g
u
l
a
r
Iss
ue
-
1
37 -
Please
c
i
t
e
t
h
is
ar
t
i
c
le
as
:
A
.
L
.
S
uárez
-C
e
t
rulo,
D
.
Q
ui
n
t
a
n
a,
A
.
C
er
v
a
n
t
es.
M
a
c
h
i
n
e
L
ear
n
i
n
g f
or
F
i
n
a
n
c
ial
Predi
ct
io
n
U
n
der
R
e
g
ime
C
h
a
n
g
e
U
si
n
g T
e
c
hn
i
c
al
A
n
al
y
sis
:
A Sy
s
t
ema
t
i
c
R
e
v
ie
w
,
I
n
t
er
n
a
t
io
n
al
J
our
n
al
o
f
I
n
t
era
ct
i
v
e
M
ul
t
imedia
a
n
d
A
r
t
i
f
i
c
ial
I
n
t
elli
g
e
n
c
e,
v
ol.
9,
n
o.
1,
pp.
137
-
148,
2024,
h
tt
p
://
d
x
.doi.or
g
/
10.9781
/
i
j
imai.2023.06.003
K
e
y
w
o
rds
C
oncep
t
Dr
if
t
,
Finance
,
M
achine
L
ea
r
ning
,
M
e
t
a
L
ea
r
ning
,
Regime
C
hange
,
S
y
s
t
ema
t
ic
L
i
t
e
r
a
t
u
r
e
Re
v
ie
w.
Abstr
a
c
t
Recen
t
c
r
ises
, r
ecessions
and
bubbles
ha
v
e
s
t
r
essed
t
he
non
-
s
t
a
t
iona
ry
na
t
u
r
e
and
t
he
p
r
esence
of
d
r
as
t
ic
s
t
r
uc
t
u
r
al
changes
in
t
he
financial
domain
. T
he
mos
t
r
ecen
t
li
t
e
r
a
t
u
r
e
sugges
t
s
t
he
use
of
con
v
en
t
ional
machine
lea
r
ning
and
s
t
a
t
is
t
ical
app
r
oaches
in
t
his
con
t
ex
t
. U
nfo
r
t
una
t
el
y,
se
v
e
r
al
of
t
hese
t
echniques
a
r
e
unable
o
r
slo
w
t
o
adap
t
t
o
changes
in
t
he
p
r
ice
-
gene
r
a
t
ion
p
r
ocess
. T
his
s
t
ud
y
aims
t
o
su
rv
e
y
t
he
r
ele
v
an
t
li
t
e
r
a
t
u
r
e
on
M
achine
L
ea
r
ning
fo
r
financial
p
r
edic
t
ion
unde
r r
egime
change
emplo
y
ing
a
s
y
s
t
ema
t
ic
app
r
oach
.
It
r
e
v
ie
w
s
k
e
y
pape
r
s
w
i
t
h
a
special
emphasis
on
t
echnical
anal
y
sis
. T
he
s
t
ud
y
discusses
t
he
g
r
o
w
ing
numbe
r
of
con
t
r
ibu
t
ions
t
ha
t
a
r
e
b
r
idging
t
he
gap
be
t
w
een
t
w
o
sepa
r
a
t
e
communi
t
ies
,
one
focused
on
da
t
a
s
t
r
eam
lea
r
ning
and
t
he
o
t
he
r
on
economic
r
esea
r
ch
. H
o
w
e
v
e
r,
i
t
also
ma
k
es
appa
r
en
t
t
ha
t
w
e
a
r
e
s
t
ill
in
an
ea
r
l
y
s
t
age
.
T
he
r
ange
of
machine
lea
r
ning
algo
r
i
t
hms
t
ha
t
ha
v
e
been
t
es
t
ed
in
t
his
domain
is
v
e
ry w
ide
,
bu
t
t
he
r
esul
t
s
of
t
he
s
t
ud
y
do
no
t
sugges
t
t
ha
t
cu
rr
en
t
l
y
t
he
r
e
is
a
specific
t
echnique
t
ha
t
is
clea
r
l
y
dominan
t
.
D
O
I:
1
0
.
9
7
8
1
/
i
j
i
m
a
i
.
2
0
2
3.
0
6
.
00
3
M
achine
L
ea
r
ning
fo
r
Financial
P
r
edic
t
ion
U
nde
r
Regime
C
hange
U
sing
T
echnical
A
nal
y
sis
: A
S
y
s
t
ema
t
ic
Re
v
ie
w
A
nd
r
és
L.
Suá
r
ez
-C
e
t
r
ulo
1
, D
a
v
id
Quin
t
ana
2
, A
lejand
r
o
C
e
rv
an
t
es
3
*
1
I
r
eland’s
C
en
t
r
e
fo
r A
pplied
A
I
(
C
e
ADA
R)
, U
ni
v
e
r
si
t
y C
olle
g
e
D
ublin
(
I
r
eland)
2
D
epa
r
t
men
t
of
C
ompu
t
e
r
Science
and
E
n
g
inee
r
in
g, U
ni
v
e
r
sidad
C
a
r
los
III
de
M
ad
r
id
, Av
da
. U
ni
v
e
r
sidad
30
,
28911
L
e
g
anes
(Spain)
3
E
scuela
Supe
r
io
r
de
I
n
g
enie
r
ía
y T
ecnolo
g
ía
, U
ni
v
e
r
sidad
I
n
t
e
r
nacional
de
L
a
Rioja
(
UN
I
R)
, L
o
gr
oño
(Spain)
* C
o
rr
espondin
g
au
t
ho
r:
alejand
r
o
.
ce
rv
an
t
es
r
o
v
i
r
a
@
uni
r.
ne
t
R
e
c
e
i
v
e
d
2
6
M
ay
2
0
22
| Acc
e
p
te
d
2
8
Apr
i
l
2
0
2
3 | Pub
l
i
sh
e
d
2
3
J
un
e
2
0
2
3
I
.
I
ntroduction
F
in
a
nci
a
l
ma
rk
e
t
s
can
be
desc
r
ibed
as
an
e
v
olu
t
iona
ry
and
nonlinea
r
d
y
namical
complex
s
y
s
t
em
[
1
]
,
[
2
]
.
Fo
r
ecas
t
ing
in
t
he
financial
domain
has
t
r
adi
t
ionall
y
been
pe
r
fo
r
med
unde
r
t
he
assump
t
ion
t
ha
t
t
he
unde
r
l
y
ing
da
t
a
has
been
c
r
ea
t
ed
b
y
a
linea
r
p
r
ocess
[
3
]
. A
no
t
he
r
line
of
w
o
rk
t
o
ma
k
e
financial
p
r
edic
t
ions
is
t
o
use
machine
lea
r
ning
(
ML
)
. T
hese
algo
r
i
t
hms
ha
v
e
su
r
p
r
ised
financial
expe
r
t
s
[
4
]
[
6
]
because
of
t
hei
r
success
in
mapping
nonlinea
r r
ela
t
ionships
w
i
t
hou
t
p
r
io
r k
no
w
ledge
[
7
]
. D
eep
lea
r
ning
algo
r
i
t
hms
(neu
r
al
ne
t
w
o
rk
s)
and
ensembles
ha
v
e
been
some
of
t
he
t
echniques
ob
t
aining
t
he
bes
t
r
esul
t
s
fo
r
s
t
oc
k
t
r
end
p
r
edic
t
ion
[
8
]
[
13
]
.
D
iffe
r
en
t
c
r
ises
, r
ecessions
and
bubbles
,
such
as
t
he
C
OV
I
D-
19
pandemic
,
o
r v
ola
t
ile
mid
-
t
e
r
m
t
r
ends
in
c
ry
p
t
o
ma
rk
e
t
s
,
ha
v
e
made
appa
r
en
t
t
he
non
-
s
t
a
t
iona
ry
na
t
u
r
e
and
t
he
p
r
esence
of
d
r
as
t
ic
s
t
r
uc
t
u
r
al
changes
in
financial
ma
rk
e
t
s
[
14
]
. D
u
r
ing
t
hese
pe
r
iods
,
mean
r
e
t
u
r
ns
, v
ola
t
ili
t
y
and
co
rr
ela
t
ions
among
asse
t
s
t
end
t
o
change
quic
k
l
y
[
15
]
. T
his
has
b
r
ough
t
a
tt
en
t
ion
t
o
t
he
p
r
oblem
of
concep
t
d
r
if
t
[
16
]
in
compu
t
a
t
ional
finance
[
17
]
. M
an
y r
ecen
t
r
esea
r
ch
w
o
rk
s
poin
t
ou
t
t
ha
t
financial
asse
t
s
o
r
companies
p
r
esen
t
diffe
r
en
t
s
t
a
t
es
t
ha
t
ma
y
r
epea
t
o
r
no
t
o
v
e
r
t
ime
o
r
e
v
ol
v
e
due
t
o
infla
t
ion
,
defla
t
ion
,
o
r
changes
in
suppl
y
and
demand
[
18
]
[
24
]
.
I
n
finance
,
a
change
in
t
he
collec
t
i
v
e
beha
v
iou
r
of
ma
rk
e
t
pa
r
t
icipan
t
s
and
t
hei
r r
eac
t
ions
is
called
a
r
egime
change
(R
C
)
. A
s
co
v
e
r
ed
b
y
t
he
ma
rk
ed
efficienc
y
h
y
po
t
hesis
[
25
]
, w
e
canno
t
obse
rv
e
t
he
indi
v
idual
beha
v
iou
r
of
a
t
r
ade
r
o
r
i
t
s
in
t
en
t
ions
.
I
ns
t
ead
, w
e
can
onl
y
obse
rv
e
changes
in
t
he
p
r
ice
d
y
namics
and
mac
r
o
o
r
mic
r
o
-
economic
v
a
r
iables
and
ex
t
r
apola
t
e
t
he
changes
t
ha
t
ma
k
e
t
hem
modif
y
t
hei
r
beha
v
iou
r. T
he
execu
t
ion
of
t
hese
s
t
r
a
t
egies
is
t
he
ac
t
ual
gene
r
a
t
i
v
e
p
r
ocess
of
t
he
obse
rv
ed
t
ime
se
r
ies
of
p
r
ices
o
r
t
r
ends
. T
he
es
t
ima
t
ion
of
t
he
hidden
p
r
ocesses
d
r
i
v
ing
t
he
ma
rk
e
t
in
t
o
diffe
r
en
t
r
egimes
is
of
t
en
app
r
oached
using
r
egime
-
s
w
i
t
ching
models
,
a
t
y
pe
of
t
ime
se
r
ies
model
w
he
r
e
pa
r
ame
t
e
r
s
can
ha
v
e
diffe
r
en
t
v
alues
in
diffe
r
en
t
c
y
cles
[
26
]
.
D
espi
t
e
t
he
fac
t
t
ha
t
a
r
t
ificial
in
t
elligence
has
r
ecen
t
l
y
become
a
t
r
end
and
e
v
en
a
buzz
w
o
r
d
in
man
y
indus
t
r
ies
,
t
his
has
no
t
become
t
he
main
t
r
end
y
e
t
fo
r
t
r
ading
s
y
s
t
ems
. T
his
is
mainl
y
due
t
o
t
he
high
complexi
t
y
and
ha
r
d
explainabili
t
y
of
t
hese
models
[
27
]
,
being
t
he
second
a
mus
t
fo
r
s
t
a
k
eholde
r
s
and
decision
-
ma
k
e
r
s
in
t
his
sec
t
o
r
[
28
]
.
I
ns
t
ead
,
t
r
ade
r
s
t
end
t
o
iden
t
if
y
di
r
ec
t
ional
changes
in
t
he
ma
rk
e
t
s
t
a
t
e
using
diffe
r
en
t
popula
r
indica
t
o
r
s
t
ailo
r
ed
acco
r
ding
t
o
t
hei
r
needs
.
Tr
adi
t
ionall
y,
t
he
li
t
e
r
a
t
u
r
e
has
used
s
t
a
t
ic
me
t
hods
t
o
in
t
e
r
p
r
e
t
pa
tt
e
r
ns
based
on
t
he
meaning
of
t
hese
indica
t
o
r
s
and
t
hei
r
his
t
o
r
ical
co
rr
ela
t
ion
t
o
fu
t
u
r
e
p
r
ices
. H
o
w
e
v
e
r,
t
his
co
rr
ela
t
ion
ma
y v
a
ry
o
v
e
r
I
ntern
a
t
i
on
a
l
J
ourn
a
l
of
I
nter
ac
t
i
v
e
M
u
l
t
i
med
i
a a
nd
Art
i
f
i
c
i
a
l
I
nte
lli
g
en
c
e,
Vo
l
.
9
,
N
º
1
-
1
3
8
-
t
ime
.
Beha
v
iou
r
al
shif
t
s
of
in
v
es
t
o
r
s
changing
con
t
inuousl
y w
i
t
h
a
hidden
con
t
ex
t
can
also
be
obse
rv
ed
t
h
r
ough
t
he
change
in
sell
v
e
r
sus
bu
y v
olumes
,
in
diffe
r
ences
be
t
w
een
local
minima
and
local
maxima
o
v
e
r
t
ime
,
and
t
h
r
ough
diffe
r
en
t
mo
v
ing
a
v
e
r
ages
a
t
diffe
r
en
t
t
ime
f
r
ames
depending
on
t
he
g
r
anula
r
de
t
ail
obse
rv
ed
(f
r
equencies)
a
t
in
t
r
ada
y,
dail
y
o
r w
ee
k
l
y
le
v
els
. C
hanges
in
financial
ma
rk
e
t
s
challenge
t
r
ade
r
s
and
in
v
es
t
o
r
s
,
as
mos
t
of
t
hei
r
models
r
el
y
on
p
r
e
v
ious
pa
tt
e
r
ns
. H
ence
,
a
w
a
y
t
o
r
ecognise
t
hese
changes
p
r
o
v
ides
a
compe
t
i
t
i
v
e
ad
v
an
t
age
since
i
t
allo
w
s
changes
in
t
r
ading
s
t
r
a
t
egies
ahead
of
o
t
he
r
in
v
es
t
o
r
s
[
29
]
. D
e
t
ec
t
ing
concep
t
shif
t
s
also
helps
lo
w
e
r
t
he
r
is
k
s
of
financial
exposu
r
e
in
high
-
f
r
equenc
y
t
r
ading
(
H
F
T
)
.
T
he
digi
t
alisa
t
ion
of
t
he
financial
indus
t
ry
has
r
esul
t
ed
in
a
g
r
o
w
ing
amoun
t
of
da
t
a
t
ha
t
is
a
v
ailable
fo
r
decision
-
ma
k
ing
. T
his
,
t
oge
t
he
r
w
i
t
h
t
he
inc
r
easing
amoun
t
of
compu
t
a
t
ional
r
esou
r
ces
,
has
accele
r
a
t
ed
t
he
adop
t
ion
of
a
w
hole
r
ange
of
machine
-
lea
r
ning
-
based
solu
t
ions
.
A
mong
t
he
sui
t
e
of
ins
t
r
umen
t
s
a
v
ailable
t
o
deal
w
i
t
h
r
egime
changes
,
online
inc
r
emen
t
al
ML
algo
r
i
t
hms
seem
especiall
y
app
r
op
r
ia
t
e
. A
mong
t
he
ad
v
an
t
ages
t
ha
t
t
he
y
offe
r, w
e
can
men
t
ion
t
he
fac
t
t
he
y
can
handle
non
-
s
t
a
t
iona
r
i
t
ies
,
shif
t
s
,
and
d
r
if
t
s
in
p
r
ice
gene
r
a
t
ion
p
r
ocesses
.
A
no
t
he
r
aspec
t
t
ha
t
ma
k
es
t
hem
a
good
fi
t
fo
r
t
his
con
t
ex
t
is
t
he
fac
t
t
ha
t
t
he
y
a
r
e
scalable
fo
r
con
t
inuous
lea
r
ning
scena
r
ios
[
30
]
.
One
migh
t
conside
r
t
w
o
main
scena
r
ios
r
ega
r
ding
t
he
na
t
u
r
e
of
s
t
r
uc
t
u
r
al
change
. T
he
fi
r
s
t
possibili
t
y
is
t
he
exis
t
ence
of
r
ecu
rr
ence
,
t
ha
t
is
,
t
he
idea
t
ha
t
t
he
s
y
s
t
em
migh
t
t
r
ansi
t
ion
bac
k
t
o
a
p
r
e
v
ious
p
r
ice
gene
r
a
t
ion
p
r
ocess
.
Fo
r
ins
t
ance
,
t
he
r
e
migh
t
be
a
specific
ma
rk
e
t
s
t
a
t
e
fo
r
ma
rk
e
t
openings
a
t
in
t
r
ada
y
f
r
equencies
and
ano
t
he
r
fo
r
financial
bubbles
t
ha
t
migh
t
be
obse
rv
able
a
t
lo
w
e
r
f
r
equencies
.
T
he
al
t
e
r
na
t
i
v
e
assumes
t
ha
t
an
y
d
r
if
t
r
esul
t
s
in
a
t
r
ansi
t
ion
t
o
a
ne
w
p
r
ocess
. A
s
w
e
w
ill
discuss
in
de
t
ail
, w
hile
t
he
r
e
is
a
r
ele
v
an
t
numbe
r
of
published
s
t
udies
on
machine
lea
r
ning
fo
r
da
t
a
s
t
r
eams
t
ha
t
pa
y
a
tt
en
t
ion
t
o
non
-
s
t
a
t
iona
r
i
t
y
[
31
]
[
35
]
,
t
he
li
t
e
r
a
t
u
r
e
on
financial
applica
t
ions
of
t
hese
algo
r
i
t
hms
,
especiall
y
t
ha
t
focused
on
r
ecu
rr
ing
concep
t
d
r
if
t
s
,
is
mo
r
e
limi
t
ed
[
17
]
,
[
36
]
[
41
]
.
W
e
mus
t
also
poin
t
ou
t
t
ha
t
t
he
p
r
edic
t
ion
of
fu
t
u
r
e
financial
t
r
ends
can
be
t
ac
k
led
using
fundamen
t
al
o
r
t
echnical
anal
y
sis
. D
espi
t
e
some
con
t
r
o
v
e
r
s
y r
ega
r
ding
i
t
s
po
t
en
t
ial
[
25
]
,
[
42
]
,
t
he
la
tt
e
r
is
v
e
ry
p
r
e
v
alen
t
in
sho
r
t
-
t
e
r
m
t
r
ading
[
43
]
,
hence
t
he
focus
on
t
his
app
r
oach
.
H
a
v
ing
said
t
ha
t
,
t
he
r
e
a
r
e
also
r
ele
v
an
t
pape
r
s
in
t
he
fi
r
s
t
ca
t
ego
ry,
li
k
e
t
he
con
t
r
ibu
t
ions
of
G
e
v
a
and
Z
aha
v
i
[
44
]
,
on
t
he
sho
r
t
-
t
e
r
m
,
in
t
r
ada
y
and
high
-
f
r
equenc
y
fo
r
ecas
t
using
ne
w
s
da
t
a
and
t
he
s
t
ud
y
of
D
og
r
a
e
t
al
.
[
4
]
,
anal
y
sing
t
he
impac
t
of
r
ecen
t
ne
w
s
on
s
t
oc
k
p
r
ice
t
r
ends
and
challenges
such
as
class
imbalance
. M
o
r
e
r
ecen
t
l
y, C
hen
e
t
al
.
[
45
]
h
y
b
r
idised
bo
t
h
app
r
oaches
in
a
s
t
ud
y
t
ha
t
combines
bo
t
h
sen
t
imen
t
anal
y
sis
and
t
echnical
indica
t
o
r
s
.
W
i
t
h
t
his
su
rv
e
y, w
e
t
ry
t
o
b
r
ing
t
o
t
he
academic
communi
t
y
’s
a
tt
en
t
ion
ho
w ML
is
being
used
t
o
deal
w
i
t
h
s
t
r
uc
t
u
r
al
change
in
financial
ma
rk
e
t
s
.
Ou
r
goal
is
t
o
iden
t
if
y
di
r
ec
t
ions
on
le
v
e
r
aging
t
he
benefi
t
s
of
mode
r
n
algo
r
i
t
hms
t
ha
t
w
o
rk w
i
t
h
diffe
r
en
t
scena
r
ios
and
deal
w
i
t
h
an
y
changes
t
ha
t
ma
y
a
r
ise
in
r
eal
-
t
ime
.
T
he
use
of
t
hese
app
r
oaches
ma
y
help
t
o
find
s
t
r
a
t
egies
t
o
imp
r
o
v
e
p
r
edic
t
ion
accu
r
ac
y
du
r
ing
t
imes
of
change
,
limi
t
ing
t
he
need
fo
r
cons
t
an
t
model
r
e
t
r
aining
. H
ence
,
some
of
t
hese
t
echniques
efficien
t
l
y
inc
r
ease
t
he
po
t
en
t
ial
p
r
ofi
t
s
,
a
v
oiding
t
he
compu
t
a
t
ional
bu
r
den
and
benefi
t
ing
f
r
om
al
w
a
y
s
ha
v
ing
an
up
-
t
o
-
da
t
e
model
.
T
he
r
es
t
of
t
he
documen
t
is
s
t
r
uc
t
u
r
ed
as
follo
w
s
:
Sec
t
ion
II
co
v
e
r
s
t
he
me
t
hodolog
y
and
r
esea
r
ch
ques
t
ions
used
in
t
his
s
y
s
t
ema
t
ic
r
e
v
ie
w;
Sec
t
ion
I
V
w
ill
discuss
t
he
ou
t
comes
of
each
of
t
he
r
esea
r
ch
ques
t
ions
. T
he
r
e
,
Subsec
t
ion
A w
ill
in
t
r
oduce
t
he
t
opic
of
r
egime
change
in
financial
se
r
ies
,
and
Sec
t
ion
B
w
ill
discuss
t
he
co
r
e
li
t
e
r
a
t
u
r
e
on
machine
lea
r
ning
fo
r
financial
p
r
edic
t
ion
unde
r r
egime
change
.
T
he
final
t
w
o
sec
t
ions
w
ill
be
r
ese
rv
ed
fo
r
a
summa
ry
of
r
esul
t
s
,
main
conclusions
,
and
fu
t
u
r
e
r
esea
r
ch
lines
.
II
.
Re
s
e
a
rch M
e
thodology
A
.
Mot
iv
at
i
on
and
Object
iv
e
s
Financial
t
ime
se
r
ies
a
r
e
of
t
en
subjec
t
t
o
s
t
r
uc
t
u
r
al
change
.
E
v
en
t
hough
machine
lea
r
ning
offe
r
s
majo
r
ad
v
an
t
ages
in
t
his
con
t
ex
t
,
t
he
li
t
e
r
a
t
u
r
e
on
t
he
t
opic
is
limi
t
ed
and
spa
r
se
. T
he
r
e
seem
t
o
be
diffe
r
en
t
r
esea
r
ch
communi
t
ies
focused
on
diffe
r
en
t
aspec
t
s
of
t
he
p
r
oblem
,
and
i
t
is
ha
r
d
t
o
k
eep
t
r
ac
k
of
t
he
main
con
t
r
ibu
t
ions
and
ins
t
r
umen
t
s
used
t
o
t
ac
k
le
t
he
p
r
oblem
.
T
he
li
t
e
r
a
t
u
r
e
p
r
esen
t
s
a
lac
k
of
s
t
udies
on
p
r
edic
t
ion
unde
r r
egime
changes
based
on
t
echnical
anal
y
sis
using
machine
lea
r
ning
. T
his
is
unfo
r
t
una
t
e
,
as
t
he
r
e
is
a
lo
t
t
o
be
gained
in
t
e
r
ms
of
efficienc
y
and
pe
r
fo
r
mance
. W
i
t
hin
t
his
field
, w
e
find
v
e
ry
p
r
omising
ideas
.
Fo
r
ins
t
ance
,
t
he
p
r
oblem
can
be
f
r
amed
using
t
he
da
t
a
s
t
r
eam
lea
r
ning
t
opic
of
concep
t
d
r
if
t
. A
significan
t
numbe
r
of
con
t
r
ibu
t
ions
t
o
t
his
ne
w
concep
t
ha
v
e
no
t
been
explici
t
l
y
applied
t
o
finance
y
e
t
,
and
t
he
y
a
r
e
no
t
w
idel
y k
no
w
n
. T
he
y
ha
v
e
no
t
been
w
idel
y
p
r
esen
t
in
machine
lea
r
ning
su
rv
e
y
s
ou
t
side
t
he
da
t
a
s
t
r
eam
lea
r
ning
niche
a
r
ea
.
Explo
r
ing
p
r
e
v
ious
r
esea
r
ch
sho
w
ed
t
ha
t
a
comp
r
ehensi
v
e
r
e
v
ie
w
does
no
t
exis
t
on
t
hese
t
opics
. T
he
r
efo
r
e
,
t
his
s
t
ud
y w
ill
help
r
eade
r
s
unde
r
s
t
and
t
he
cu
rr
en
t
s
t
a
t
e
of
t
he
a
r
t
,
b
r
idge
t
he
gap
among
r
esea
r
ch
fields
,
and
add
r
ess
p
r
omising
fu
t
u
r
e
lines
of
r
esea
r
ch
in
t
his
domain
.
B
.
Re
s
earc
h
Met
h
od
I
n
o
r
de
r
t
o
p
r
o
v
ide
an
o
v
e
rv
ie
w
of
t
he
s
t
a
t
e
of
t
he
ques
t
ion
,
ou
r
r
esea
r
ch
has
follo
w
ed
K
i
t
chenham
and
C
ha
r
t
e
r
s’
guidelines
on
S
y
s
t
ema
t
ic
L
i
t
e
r
a
t
u
r
e
Re
v
ie
w
(S
L
R)
[
46
]
,
[
47
]
.
A
s
y
s
t
ema
t
ic
r
e
v
ie
w
is
defined
as
an
o
r
ganised
w
a
y
t
o
s
y
n
t
hesise
exis
t
ing
w
o
rk
fai
r
l
y. A
n
S
L
R
is
a
means
t
o
iden
t
if
y,
e
v
alua
t
e
and
in
t
e
r
p
r
e
t
t
he
a
v
ailable
r
esea
r
ch
w
o
rk
s
r
ele
v
an
t
t
o
a
defini
t
e
r
esea
r
ch
ques
t
ion
,
t
opic
a
r
ea
,
o
r
phenomenon
of
in
t
e
r
es
t
. A
f
t
e
r r
e
v
ising
t
he
li
t
e
r
a
t
u
r
e
fo
r
simila
r r
esea
r
ch
objec
t
i
v
es
,
i
t
can
be
iden
t
ified
t
ha
t
t
he
r
e
is
no
p
r
e
v
iousl
y
published
sea
r
ch
on
a
t
opic
.
C
.
P
l
ann
i
n
g
T
he
s
t
ud
y
aims
t
o
summa
r
ise
t
he
cu
rr
en
t
s
t
a
t
us
of
p
r
edic
t
ing
financial
t
ime
se
r
ies
in
t
he
financial
li
t
e
r
a
t
u
r
e
du
r
ing
beha
v
iou
r
al
o
r
r
egime
changes
in
ma
rk
e
t
s
. K
i
t
chenham
and
C
ha
r
t
e
r
s’
S
L
R
p
r
o
t
ocol
w
as
adap
t
ed
t
o
desc
r
ibe
t
he
plan
fo
r
t
he
r
e
v
ie
w.
T
he
p
r
o
t
ocol
comp
r
ises
r
esea
r
ch
bac
k
g
r
ound
and
ques
t
ions
,
sea
r
ch
s
t
r
a
t
eg
y,
s
t
ud
y
selec
t
ion
c
r
i
t
e
r
ia
and
p
r
ocedu
r
es
,
da
t
a
ex
t
r
ac
t
ion
,
and
da
t
a
s
y
n
t
hesis
s
t
r
a
t
egies
t
o
gua
r
an
t
ee
t
ha
t
t
he
in
v
es
t
iga
t
ion
is
unde
r
t
a
k
en
as
in
t
ended
and
r
educe
t
he
li
k
elihood
of
bias
in
t
he
s
t
ud
y.
I
n
t
his
p
r
o
t
ocol
,
t
he
en
t
i
r
e
in
v
es
t
iga
t
ion
plan
w
as
no
t
decided
f
r
om
t
he
beginning
.
I
ns
t
ead
,
t
his
and
t
he
r
esul
t
s
p
r
oduced
w
e
r
e
r
eco
r
ded
as
t
he
s
t
ud
y
p
r
og
r
essed
.
D
.
Re
s
earc
h
Q
u
e
s
t
i
on
s
T
his
pape
r
has
t
he
follo
w
ing
t
w
o
r
esea
r
ch
ques
t
ions
:
Q1
W
ha
t
a
r
e
t
he
diffe
r
en
t
r
esea
r
ch
a
r
eas
fo
r
p
r
edic
t
ing
unde
r r
egime
changes
in
t
he
financial
li
t
e
r
a
t
u
r
e?
and
Q2
W
ha
t
a
r
e
t
he
mos
t
commonl
y
used
machine
lea
r
ning
t
echniques
applied
t
o
anal
y
sing
r
egime
changes?
T
he
r
esul
t
s
expec
t
ed
a
t
t
he
end
of
t
he
s
y
s
t
ema
t
ic
r
e
v
ie
w w
e
r
e
t
o
see
w
ha
t
r
esea
r
ch
o
r
su
rv
e
y
s
had
been
applied
o
r
p
r
oduced
on
t
he
t
opic
so
fa
r
and
t
o
iden
t
if
y
t
he
implica
t
ions
of
using
machine
lea
r
ning
t
o
handle
beha
v
iou
r
al
changes
in
financial
ma
rk
e
t
s
in
t
he
scien
t
ific
li
t
e
r
a
t
u
r
e
.
E
.
Searc
h
Strate
g
y
and
P
roce
ss
T
he
sea
r
ch
s
t
r
a
t
eg
y
included
:
i)
sea
r
ch
r
esou
r
ces
and
ii)
a
sea
r
ch
p
r
ocess
.
Each
one
of
t
hem
is
de
t
ailed
in
t
he
follo
w
ing
subsec
t
ions
.
R
e
g
u
l
a
r
Iss
ue
-
1
3
9
-
1.
Searc
h
Re
s
o
u
rce
s
T
his
s
t
ud
y w
as
planned
t
o
find
all
t
he
li
t
e
r
a
t
u
r
e
a
v
ailable
abou
t
machine
lea
r
ning
fo
r
fo
r
ecas
t
ing
unde
r r
egime
changes
in
finance
.
T
he
sou
r
ces
used
fo
r
t
he
s
y
s
t
ema
t
ic
r
e
v
ie
w w
e
r
e
:
I
EEE
D
igi
t
al
L
ib
r
a
ry
(h
tt
p
:
//ieeexplo
r
e
.
ieee
.
o
r
g)
;
Science
D
i
r
ec
t
,
on
t
he
subjec
t
of
C
ompu
t
e
r
Science
(h
tt
ps
:
//
www.
sciencedi
r
ec
t
.
com/)
;
ACM D
igi
t
al
L
ib
r
a
ry
(h
tt
p
:
//po
r
t
al
.
acm
.
o
r
g)
;
T
a
y
lo
r
&
F
r
ancis
Jou
r
nals
(h
tt
p
:
//
www.
t
andfonline
.
com)
;
W
ile
y
Online
L
ib
r
a
ry
(h
tt
p
:
//
www.w
ile
y.
com/)
;
Sp
r
inge
rL
in
k
(h
tt
p
:
//lin
k.
sp
r
inge
r.
com)
;
and
addi
t
ionall
y
G
oogle
Schola
r w
as
explo
r
ed
as
g
r
e
y
li
t
e
r
a
t
u
r
e
(h
tt
ps
:
//schola
r.
google
.
com/)
.
2.
Searc
h
P
roce
ss
T
he
o
v
e
r
all
sea
r
ch
p
r
ocess
is
depic
t
ed
in
Fig
.
1
and
is
explained
in
t
he
follo
w
ing
sec
t
ion
.
I
d
e
n
ti
f
i
ca
ti
on
o
f
s
t
ud
ie
s
vi
a
da
t
abas
e
s
I
d
e
n
tifi
ca
ti
on
S
c
ree
n
i
n
g
I
nc
l
ud
e
d
R
e
c
o
rds r
e
m
o
v
e
d
b
e
f
o
r
e
sc
r
ee
n
i
ng
:
R
e
c
o
rds
i
d
e
nt
ii
e
d fr
o
m
:
D
up
li
cat
e
r
e
c
o
rd r
e
m
o
v
e
d
D
atabas
e
s (n =
7
)
(n =
20
)
R
e
gi
st
e
rs (n =
0
)
R
e
c
o
rds r
e
m
o
v
e
d f
o
r n
o
t
b
e
i
n
g
ava
il
ab
l
e
o
r n
o
t hav
i
n
g
a subscr
i
pt
i
o
n (n =
8
)
R
e
c
o
rds scr
ee
n
e
d
R
e
c
o
rds
e
xc
l
ud
e
d bas
e
d
o
n t
i
t
l
e
(n =
643
)(n =
217
)
R
e
p
o
rts s
o
u
g
ht f
o
r r
e
tr
e
i
e
va
l
(n =
426
)
R
e
p
o
rts n
o
t r
e
tr
i
e
v
e
d a�
e
r
r
e
ad
i
n
g
abstract and
k
e
y
w
o
rds
(n =
203
)
R
e
p
o
rts
e
xc
l
ud
e
d bas
e
d
o
n fu
ll
t
e
xt
:
S
tud
i
e
s
o
n
ly
f
o
cus
i
n
g
o
n ML
m
e
th
o
ds (n =
61
)
Fi
nanc
i
a
l
r
e
gi
m
e
chan
g
e
w
o
r
k
s n
o
t f
o
cus
e
d
i
n
t
e
chn
i
qu
e
s (n =
22
)
R
e
p
o
rts ass
e
ss
e
d f
o
r
e
ligi
b
ili
t
y
(n =
223
)
S
tud
i
e
s
i
nc
l
ud
e
d
i
n r
e
v
i
e
w
(n =
140
)
Fi
g.
1
.
Flo
w
of
info
r
ma
t
ion
t
h
r
ou
g
h
t
he
di
ff
e
r
en
t
phases
of
t
he
r
e
v
ie
w
usin
g
a
PR
I
S
MA
dia
gr
am
[
48
]
.
T
he
s
t
a
r
t
ing
poin
t
w
as
choosing
a
se
t
of
r
ele
v
an
t
k
e
yw
o
r
ds
. T
he
y
w
e
r
e
:
re
gim
e
chan
g
e
,
re
gim
e
-
s
wi
tch
i
n
g m
ode
l, m
ach
i
ne
l
earn
i
n
g,
chan
g
e
detect
i
on
and
s
toc
k
trend
f
oreca
s
t
i
n
g. T
he
sea
r
ch
w
as
t
hen
r
un
on
t
he
al
r
ead
y
men
t
ioned
da
t
abases
in
M
a
r
ch
2022
, r
e
t
u
r
ning
643
w
o
rk
s
in
t
o
t
al
in
a
t
ime
r
ange
,
including
t
he
y
ea
r
s
1970
t
o
2022
.
I
rr
ele
v
an
t
and
duplica
t
e
publica
t
ions
w
e
r
e
r
emo
v
ed
,
and
223
unique
r
esea
r
ch
w
o
rk
s
r
emained
. A
t
t
ha
t
poin
t
,
publica
t
ions
w
e
r
e
r
e
v
ie
w
ed
based
on
t
i
t
les
,
abs
t
r
ac
t
s
,
conclusions
, r
efe
r
ences
and
k
e
yw
o
r
ds
and
t
hen
w
e
r
e
classified
in
t
o
t
h
r
ee
diffe
r
en
t
t
y
pes
:
Rele
v
an
t
w
o
rk
s
:
t
hese
should
sa
t
isf
y
one
of
t
he
t
w
o
inclusion
c
r
i
t
e
r
ia
co
v
e
r
ed
la
t
e
r
in
t
his
subsec
t
ion
;
P
r
ocess
assessmen
t
w
o
rk
s
:
if
t
he
publica
t
ion
is
r
ela
t
ed
t
o
t
he
financial
domain
o
r
concep
t
d
r
if
t
li
t
e
r
a
t
u
r
e
and
is
r
ele
v
an
t
.
Excluded
w
o
rk
s
: w
o
rk
s
no
t
r
ele
v
an
t
t
o
t
he
t
opic
.
W
hen
t
he
r
e
w
as
doub
t
abou
t
t
he
classifica
t
ion
of
a
r
esea
r
ch
w
o
rk
piece
,
i
t
w
as
included
in
t
he
r
ele
v
an
t
g
r
oup
,
lea
v
ing
t
he
possibili
t
y
of
disca
r
ding
i
t
du
r
ing
t
he
nex
t
s
t
age
, w
hen
t
he
full
-
t
ex
t
v
e
r
sions
w
e
r
e
r
e
v
ie
w
ed
. T
hi
r
d
,
each
full
a
r
t
icle
w
as
r
e
t
r
ie
v
ed
and
r
ead
t
o
v
e
r
if
y
i
t
s
inclusion
o
r
exclusion
. T
he
r
eason
fo
r
exclusion
o
r
inclusion
in
t
his
t
hi
r
d
s
t
age
w
as
documen
t
ed
.
Fou
r
t
h
,
t
o
chec
k
t
he
consis
t
enc
y
of
t
he
inclusion/exclusion
decisions
,
a
t
es
t
-r
e
t
es
t
app
r
oach
and
r
e
-
e
v
alua
t
ion
of
a
r
andom
sample
of
t
he
p
r
ima
ry
s
t
udies
w
e
r
e
made
.
D
ocumen
t
s
w
e
r
e
k
ep
t
w
hen
t
he
y
sa
t
isfied
a
t
leas
t
one
of
t
he
c
r
i
t
e
r
ia
belo
w:
T
he
w
o
rk w
as
explici
t
l
y r
ela
t
ed
t
o
r
egime
changes
o
r
s
t
r
uc
t
u
r
al
b
r
ea
k
s
in
non
-
s
t
a
t
iona
ry
da
t
a
.
T
he
w
o
rk w
as
r
ele
v
an
t
t
o
machine
lea
r
ning
fo
r
ecas
t
ing
in
domains
w
i
t
h
complex
d
y
namics
and
non
-
s
t
a
t
iona
r
i
t
ies
in
t
he
financial
field
.
T
he
au
t
ho
r
s
r
e
v
ie
w
ed
all
223
r
esea
r
ch
w
o
rk
s
and
pu
t
t
hem
in
t
o
t
hese
diffe
r
en
t
g
r
oups
acco
r
ding
t
o
t
he
p
r
e
v
iousl
y
men
t
ioned
c
r
i
t
e
r
ia
.
T
his
lis
t
w
as
r
e
v
ie
w
ed
t
o
chec
k
fo
r
inconsis
t
encies
. T
he
r
esul
t
of
t
his
s
t
age
w
as
t
ha
t
140
publica
t
ions
w
e
r
e
classified
as
r
ele
v
an
t
.
T
he
r
e
is
a
r
is
k
t
ha
t
some
r
ele
v
an
t
w
o
rk
s
ha
v
e
been
missed
.
T
he
r
efo
r
e
,
t
his
s
t
ud
y
canno
t
gua
r
an
t
ee
comple
t
eness
. H
o
w
e
v
e
r,
i
t
can
s
t
ill
be
t
r
us
t
ed
t
o
gi
v
e
a
good
o
v
e
rv
ie
w
of
t
he
r
ele
v
an
t
li
t
e
r
a
t
u
r
e
on
p
r
ice
fo
r
ecas
t
ing
in
t
he
financial
domain
unde
r
s
t
r
uc
t
u
r
al
b
r
ea
k
s
.
3.
D
ata
E
x
tract
i
on
T
he
da
t
a
ex
t
r
ac
t
ed
f
r
om
each
publica
t
ion
w
as
documen
t
ed
and
k
ep
t
in
a
r
efe
r
ence
manage
r. A
f
t
e
r
t
he
iden
t
ifica
t
ion
of
t
he
publica
t
ions
,
t
he
follo
w
ing
w
as
ex
t
r
ac
t
ed
:
Sou
r
ce
(jou
r
nal
,
boo
k,
confe
r
ence
o
r
s
t
r
ic
t
l
y r
ele
v
an
t
t
echnical
o
r
w
hi
t
e
pape
r
)
;
T
i
t
le
;
Publica
t
ion
y
ea
r;
A
u
t
ho
r
s
;
C
lassifica
t
ion
acco
r
ding
t
o
t
opics
;
Summa
ry
of
t
he
r
esea
r
ch
,
including
w
hich
ques
t
ions
w
e
r
e
sol
v
ed
.
III
.
S
umm
a
ry of
Re
sults
I
n
o
r
de
r
t
o
anal
y
se
t
he
223
w
o
rk
s
, w
e
found
t
he
need
t
o
classif
y
t
hem
in
mo
r
e
w
a
y
s
t
han
jus
t
acco
r
ding
t
o
t
he
me
t
hodolog
y
defined
in
Sec
t
ion
II
. W
hen
needed
,
t
he
t
opics
w
e
r
e
upda
t
ed
o
r
cla
r
ified
du
r
ing
t
he
classifica
t
ion
p
r
ocess
.
Resul
t
s
of
t
he
classifica
t
ion
p
r
ocess
w
i
t
h
r
ega
r
d
t
o
t
he
r
esea
r
ch
ques
t
ions
a
r
e
de
t
ailed
in
T
able
I
.
TABLE
I
. Cl
a
ssific
a
tion of
P
a
p
e
rs With
Re
g
a
rd to th
e
Re
s
e
a
rch
Q
u
e
stions
Ques
t
ion
T
opicRele
v
an
t
S
t
udiesQuan
t
i
t
y
Q1Re
g
ime
chan
g
es
[
14
]
,
[
15
]
,
[
18
]
,
[
19
]
,
[
23
]
,
22
[
26
]
,
[
29
]
,
[
49
]
[
63
]
Q1
and
Q2
ML
in
s
t
oc
k
[
1
]
[
13
]
,
[
20
]
[
22
]
,
[
24
]
,
73
fo
r
ecas
t
in
g
[
25
]
,
[
27
]
,
[
28
]
,
[
37
]
,
[
38
]
,
[
42
]
[
45
]
,
[
64
]
[
110
]
Q2
C
oncep
t
d
r
if
t
[
16
]
,
[
17
]
,
[
30
]
[
36
]
,
45
and
online
ML
[
39
]
[
41
]
,
[
111
]
[
143
]
T
he
da
t
a
r
equi
r
ed
fo
r
anal
y
sis
w
e
r
e
ex
t
r
ac
t
ed
b
y
explo
r
ing
t
he
full
t
ex
t
of
each
r
esea
r
ch
w
o
rk. T
able
II
p
r
esen
t
s
t
he
r
esul
t
s
of
t
he
sea
r
ch
and
t
he
sou
r
ce
of
t
he
documen
t
s
. T
able
III
p
r
esen
t
s
t
he
r
esul
t
s
in
t
he
second
s
t
age
. A
s
men
t
ioned
befo
r
e
,
t
he
t
o
t
al
numbe
r
of
pape
r
s
r
emaining
af
t
e
r
t
he
exclusion
p
r
ocess
w
as
140
. T
able
I
summa
r
ises
t
hei
r
classifica
t
ion
acco
r
ding
t
o
t
he
k
no
w
ledge
a
r
ea
.
I
ntern
a
t
i
on
a
l
J
ourn
a
l
of
I
nter
ac
t
i
v
e
M
u
l
t
i
med
i
a a
nd
Art
i
f
i
c
i
a
l
I
nte
lli
g
en
c
e,
Vo
l
.
9
,
N
º
1
-
1
4
0 -
T
he
r
ele
v
ance
of
r
egime
changes
o
r
s
t
r
uc
t
u
r
al
b
r
ea
k
s
in
t
he
li
t
e
r
a
t
u
r
e
of
financial
p
r
ice
fo
r
ecas
t
ing
leads
t
o
conside
r
t
w
o
majo
r
a
r
eas
:
financial
r
egime
changes
(
r
ela
t
ed
t
o
majo
r
l
y
s
t
a
t
is
t
ical
app
r
oaches
t
o
de
t
ec
t
change
poin
t
s
o
r
fo
r
ecas
t
unde
r
diffe
r
en
t
r
egimes)
and
da
t
a
s
t
r
eam
lea
r
ning
(
w
he
r
e
t
he
p
r
oblem
of
concep
t
d
r
if
t
can
be
unde
r
s
t
ood
as
a
t
y
pe
of
r
egime
change
in
t
he
ML
li
t
e
r
a
t
u
r
e)
.
TABLE
II
.
Re
sults Without
F
ilt
e
ring
D
ata
So
ur
ce
T
ota
l
Pu
b
li
cat
i
o
n
s
Science
D
i
r
ec
t
76
G
oo
g
le
Schola
r
56
Sp
r
in
g
e
r
lin
k
33
I
EEE
D
i
g
i
t
al
L
ib
r
a
ry
27
ACM
D
i
g
i
t
al
L
ib
r
a
ry
14
W
ile
y
9
T
a
y
lo
r
&
F
r
ancis
8
TABLE
III
.
Se
cond
S
t
a
g
e
Re
sults
D
ata
So
ur
ce
T
ota
l
Pu
b
li
cat
i
o
n
s
Science
D
i
r
ec
t
56
G
oo
g
le
Schola
r
34
Sp
r
in
g
e
r
lin
k
20
I
EEE
D
i
g
i
t
al
L
ib
r
a
ry
16
T
a
y
lo
r
&
F
r
ancis
5
W
ile
y
5
ACM
D
i
g
i
t
al
L
ib
r
a
ry
4
Fig
.
2
sho
w
s
ho
w,
ou
t
of
a
t
o
t
al
of
140
r
ele
v
an
t
s
t
udies
,
t
he
majo
r
i
t
y
of
t
he
w
o
rk
s
r
e
v
ie
w
ed
t
o
co
rr
espond
t
o
ML
t
echniques
applied
t
o
s
t
oc
k
fo
r
ecas
t
ing
.
Some
of
t
hese
w
o
rk
s
o
v
e
r
lap
r
egime
change
r
esea
r
ch
,
focusing
p
r
ima
r
il
y
on
p
r
obabilis
t
ic
models
t
o
classif
y
di
r
ec
t
ional
changes
and
r
ep
r
esen
t
diffe
r
en
t
r
egimes
. T
he
li
t
e
r
a
t
u
r
e
on
online
lea
r
ning
does
no
t
t
end
t
o
coincide
w
i
t
h
t
he
one
on
r
egime
changes
. H
o
w
e
v
e
r,
s
t
udies
of
online
ML
t
ac
k
le
simila
r
challenges
as
models
t
o
handle
r
egime
changes
,
such
as
ha
v
ing
up
-
t
o
-
da
t
e
models
and
r
e
-
t
r
aining
mechanisms
.
A
deepe
r
discussion
on
t
his
ma
tt
e
r w
ill
be
held
in
Sec
t
ion
I
V
.
Fi
g.
2
. T
opic
dis
t
r
ibu
t
ion
of
r
esea
r
ch
pape
r
s
af
t
e
r
f
il
t
e
r
in
g.
Fig
.
3
sho
w
s
t
he
dis
t
r
ibu
t
ion
of
pape
r
s
r
e
v
ie
w
ed
ac
r
oss
v
a
r
ious
sou
r
ces
. A
majo
r
i
t
y
of
t
he
r
esea
r
ch
w
o
rk
s
ha
v
e
been
r
e
t
r
ie
v
ed
f
r
om
high
-
impac
t
jou
r
nals
,
follo
w
ed
b
y
confe
r
ences
and
boo
k
s
. H
o
w
e
v
e
r,
since
some
of
t
he
t
opics
r
e
v
ie
w
ed
,
li
k
e
online
ML,
a
r
e
cu
rr
en
t
r
esea
r
ch
a
r
eas
,
a
r
emainde
r,
close
t
o
«
4
%
of
r
esea
r
ch
w
o
rk
s
,
belong
t
o
non
-
pee
r-r
e
v
ie
w
ed
pape
r
s
con
t
ained
b
y
open
-
access
r
eposi
t
o
r
ies
.
Finall
y, w
e
ha
v
e
ex
t
r
ac
t
ed
f
r
om
t
he
pape
r
s
classified
pape
r
s
unde
r
t
he
t
opic
"
C
oncep
t
Dr
if
t
"
t
he
ML
t
echnique
mainl
y
used
,
ei
t
he
r
as
a
ne
w
p
r
oposal
o
r
as
a
r
efe
r
ence
fo
r
compa
r
ison
. W
e
ha
v
e
g
r
ouped
t
hese
t
echniques
in
t
o
eigh
t
b
r
oad
ca
t
ego
r
ies
(Fig
.
4
and
5)
.
Fo
r
t
his
t
as
k, w
e
ha
v
e
excluded
r
e
v
ie
w
s
. T
hese
r
esul
t
s
sho
w
t
ha
t
mos
t
of
t
he
r
e
v
ie
w
ed
pape
r
s
use
t
echniques
f
r
om
fou
r
main
ca
t
ego
r
ies
:
E
v
ol
v
ing
s
y
s
t
ems
(
t
ha
t
include
E
v
ol
v
ing
clus
t
e
r
ing
,
E
v
ol
v
ing
fuzz
y r
ules
and
Fuzz
y
neu
r
o
s
y
s
t
ems)
,
Ensemble
based
s
y
s
t
ems
(usuall
y w
i
t
h
t
r
ee
-
based
componen
t
s)
,
t
r
adi
t
ional
s
y
s
t
ems
adap
t
ed
t
o
concep
t
change
(such
as
adap
t
i
v
e
decision
t
r
ees)
,
and
finall
y
Neu
r
al
Ne
t
w
o
rk
s
and
D
eep
L
ea
r
ning
. T
he
la
tt
e
r
a
r
e
mo
r
e
r
ecen
t
in
gene
r
al
,
and
t
he
r
efo
r
e
t
his
t
r
end
is
li
k
el
y
t
o
become
mo
r
e
impo
r
t
an
t
in
t
he
nea
r
fu
t
u
r
e
.
Fig
.
6
sho
w
s
t
he
e
v
olu
t
ion
of
t
hese
ca
t
ego
r
ies
.
F
r
eq
u
e
nc
y
(Lo
g
sca
l
e)
78
.
57%
10
.
00%
7
.
86%
2
.
14%
1
.
43%
r
epo
rt
J
o
urna
l
Co
nf
e
r
e
nc
eBoo
kT
e
chn
i
ca
l
Wh
i
t
e p
a
pe
r
Fi
g.
3
.
Sou
r
ce
dis
t
r
ibu
t
ion
of
r
esea
r
ch
pape
r
s
af
t
e
r
f
il
t
e
r
in
g.
25.00
%
17.86
%
12.50
%
12.50
%
8.93
%
8.93
%
7.14
%
7.14
%
En
s
e
mb
l
e
s D
e
c
i
s
i
o
nEv
o
lv
i
ng
T
r
ee
s (s
i
ngl
e
)
Fuzz
y
R
ul
e
s
N
e
ur
a
l
B
a
y
e
s
i
a
n
Ev
o
lv
i
ng
Fuzz
y
N
e
ur
o K
NN
N
e
t
wo
r
k
s a
n
d
M
a
r
k
o
v
C
lu
s
t
e
r
i
ng
S
y
s
t
e
ms a
n
d
S
V
M
F
r
e
qu
e
nc
y
Fi
g.
4
. ML
t
echniques
found
in
C
oncep
t
Dr
if
t
s
t
udies
, gr
ouped
b
y
ca
t
e
g
o
r
ies
,
coun
t
in
g
each
di
ff
e
r
en
t
t
echnique
used
in
pape
r
s
and
assi
g
nin
g
t
he
co
rr
espondin
g
ca
t
e
g
o
ry
sepa
r
a
t
el
y.
I
n
t
his
f
i
g
u
r
e
,
a
sin
g
le
pape
r
compa
r
in
g
se
v
e
r
al
al
g
o
r
i
t
hms
in
t
he
same
ca
t
e
g
o
ry
is
coun
t
ed
as
man
y
t
imes
as
al
g
o
r
i
t
hms
.
F
r
e
qu
e
nc
y
R
e
g
i
m
e
c
h
a
ng
e
s
Co
n
c
e
pt dr
i
15
.
7%
32
.
1%
52
.
1%
15.56
%
15.56
%
15.56
%
15.56
%
M
L
i
n
stoc
k
f
or
e
cast
i
ng
11.11
%
8.89
%
8.89
%
8.89
%
D
e
c
i
s
i
o
n
En
s
e
mb
l
e
s
Ev
o
lv
i
ng
N
e
ur
a
l
B
a
y
e
s
i
a
n
Ev
o
lv
i
ng
Fuzz
y
N
e
ur
o K
NN
T
r
ee
s (s
i
ngl
e
)
Fuzz
y
N
e
t
wo
r
k
s a
n
d
M
a
r
k
o
v
C
lu
s
t
e
r
i
ng
S
y
s
t
e
ms a
n
d
S
V
M
R
ul
e
s
Fi
g.
5
. Ty
pes
of
ML
found
in
C
oncep
t
Dr
if
t
s
t
udies
,
usin
g
t
he
unique
ca
t
e
g
o
r
ies
found
in
t
he
same
pape
r.
I
n
t
his
f
i
g
u
r
e
,
a
sin
g
le
pape
r
compa
r
in
g
six
me
t
hods
in
ca
t
e
g
o
ry A
and
one
me
t
hod
in
ca
t
e
g
o
ry B
is
coun
t
ed
as
onl
y
t
w
o
en
t
r
ies
(one
fo
r A
and
ano
t
he
r
fo
r B
)
.
R
e
g
u
l
a
r
Iss
ue
-
1
4
1
-
B
a
y
e
s
i
an and Mar
k
o
v
D
e
c
i
s
i
o
n
T
r
ee
s (adapt
e
d)
E
ns
e
mb
l
e
s
E
v
o
l
v
i
n
g
C
l
ust
e
r
i
n
g
E
v
o
l
v
i
n
g
F
uzz
y
R
u
l
e
s
F
uzz
y
N
e
ur
o
Sy
st
e
ms
K
NN
and
SV
M
N
e
ura
l
N
e
tw
o
r
k
s
ML t
e
chn
i
qu
e
i
n c
o
ll
e
ct
i
o
n
0
<
20052005
-
2015
>
2015
Y
e
ars
5
10
15
20
25
Fi
g.
6
. C
a
t
e
g
o
r
ies
pe
r
pe
r
iod
of
10
y
ea
r
s
based
on
da
t
a
used
in
Fi
g.
5
.
I
V. Discussion
T
his
sec
t
ion
desc
r
ibes
t
he
pape
r
s
r
e
v
ie
w
ed
in
t
his
w
o
rk.
I
n
t
his
discussion
, w
e
follo
w
t
he
schema
in
Fig
.
7
.
A
.
Re
gim
e
Ch
an
g
e
s
i
n
F
i
nanc
i
a
l
Ser
i
e
s
(Q
1
)
Ea
r
l
y
s
t
udies
f
r
om
t
he
financial
li
t
e
r
a
t
u
r
e
claim
t
ha
t
financial
ma
rk
e
t
s
a
r
e
efficien
t
[
25
]
and
,
as
a
r
esul
t
,
asse
t
p
r
ices
follo
w
a
r
andom
w
al
k
[
81
]
.
Fama
[
25
]
claimed
t
ha
t
ma
rk
e
t
s
canno
t
be
consis
t
en
t
l
y
bea
t
en
on
a
r
is
k-
adjus
t
ed
basis
and
t
ha
t
t
hei
r
p
r
ices
canno
t
be
an
t
icipa
t
ed
has
al
w
a
y
s
been
a
sou
r
ce
of
con
t
r
o
v
e
r
s
y
in
t
he
li
t
e
r
a
t
u
r
e
.
M
an
y r
esea
r
ch
w
o
rk
s
ha
v
e
poin
t
ed
t
o
diffe
r
en
t
ma
rk
e
t
s
being
p
r
edic
t
able
using
diffe
r
en
t
sou
r
ces
of
info
r
ma
t
ion
[
5
]
,
[
77
]
,
[
78
]
,
[
85
]
,
[
88
]
,
[
104
]
.
Fo
r
ecas
t
ing
in
t
he
financial
domain
can
be
cha
r
ac
t
e
r
ised
b
y
a
non
-
s
t
a
t
iona
ry
and
uns
t
r
uc
t
u
r
ed
na
t
u
r
e
and
b
y
hidden
r
ela
t
ionships
[
2
]
,
[
74
]
.
Economic
,
social
and
poli
t
ical
fac
t
o
r
s
w
i
t
hin
coun
t
r
ies
and
in
t
e
r
na
t
ional
impac
t
add
unce
r
t
ain
t
y
t
o
financial
ma
rk
e
t
s
[
66
]
,
[
70
]
,
[
79
]
,
[
93
]
,
[
94
]
,
[
100
]
,
[
106
]
. H
ence
,
ma
rk
e
t
s
can
be
conside
r
ed
an
e
v
olu
t
iona
ry
and
nonlinea
r
complex
s
y
s
t
em
[
1
]
. T
he
financial
li
t
e
r
a
t
u
r
e
has
co
v
e
r
ed
diffe
r
en
t
app
r
oaches
t
o
p
r
edic
t
ing
ma
rk
e
t
p
r
ices
using
s
t
a
t
is
t
ical
and
,
mo
r
e
r
ecen
t
l
y, A
I
-
based
me
t
hods
.
I
n
r
ecen
t
y
ea
r
s
,
diffe
r
en
t
e
v
en
t
s
li
k
e
t
he
C
OV
I
D-
19
pandemic
o
r
t
he
ban
kr
up
t
c
y
of
L
ehman
B
r
o
t
he
r
s
in
2008
ha
v
e
led
t
o
pe
r
iods
w
i
t
h
changes
in
mean
, v
ola
t
ili
t
y
and
co
rr
ela
t
ions
in
s
t
oc
k
ma
rk
e
t
r
e
t
u
r
ns
[
15
]
,
s
t
r
essing
t
he
non
-
s
t
a
t
iona
ry
na
t
u
r
e
and
t
he
exis
t
ence
of
d
r
as
t
ic
s
t
r
uc
t
u
r
al
changes
in
financial
ma
rk
e
t
s
[
18
]
,
[
20
]
[
23
]
.
I
n
t
he
financial
li
t
e
r
a
t
u
r
e
,
changes
in
t
he
p
r
ice
beha
v
iou
r
of
financial
ma
rk
e
t
s
t
ha
t
go
be
y
ond
t
hei
r
no
r
mal
p
r
ice
fluc
t
ua
t
ions
r
ecei
v
e
t
he
name
of
r
egime
changes
[
19
]
,
[
53
]
,
[
63
]
o
r
business
c
y
cles
shif
t
s
[
80
]
.
I
n
o
r
de
r
t
o
model
t
hese
r
egime
changes
,
one
of
t
he
mos
t
popula
r
t
echniques
is
t
he
r
egime
-
s
w
i
t
ching
model
[
15
]
, w
hich
w
as
fi
r
s
t
applied
b
y H
amil
t
on
[
58
]
as
a
t
echnique
t
o
deal
w
i
t
h
c
y
cles
of
diffe
r
en
t
economic
ac
t
i
v
i
t
ies
such
as
r
ecessions
and
ma
rk
e
t
expansions
.
I
n
financial
ma
rk
e
t
s
,
t
he
r
e
a
r
e
pe
r
iods
of
t
ime
w
i
t
h
diffe
r
en
t
deg
r
ees
of
efficienc
y
and
p
r
edic
t
abili
t
y. T
he
r
e
can
be
momen
t
s
w
he
r
e
,
due
t
o
t
he
ma
rk
e
t
-w
ide
sen
t
imen
t
gi
v
en
b
y
poli
t
ical
o
r
economic
ci
r
cums
t
ances
,
t
he
beha
v
iou
r
of
in
v
es
t
o
r
s
ma
y
change
t
o
w
a
r
ds
a
bea
r,
bull
,
la
t
e
r
al
ma
rk
e
t
,
and
pe
r
iods
o
r
t
ime
f
r
ames
w
i
t
h
diffe
r
en
t
le
v
els
of
v
ola
t
ili
t
y
[
80
]
.
A
t
t
he
mac
r
oeconomic
le
v
el
,
R
C
a
r
e
of
t
en
r
ela
t
ed
t
o
ab
r
up
t
b
r
ea
k
s
in
long
-
t
e
r
m
c
y
cles
li
k
e
t
he
b
r
ea
k
of
bubbles
o
r
economic
c
r
ises
[
59
]
. C
hanges
in
ma
rk
e
t
r
egimes
could
be
d
r
i
v
en
as
w
ell
b
y
in
v
es
t
o
r
expec
t
a
t
ions
[
15
]
. T
he
financial
li
t
e
r
a
t
u
r
e
iden
t
ifies
t
w
o
t
y
pes
of
r
egimes
clea
r
t
o
r
ecognise
:
s
t
ead
y
and
highl
y v
ola
t
ile
r
egimes
usuall
y
lin
k
ed
t
o
economic
g
r
o
w
t
h
o
r
defla
t
ion
pe
r
iods
, r
espec
t
i
v
el
y.
T
his
is
illus
t
r
a
t
ed
in
Fig
.
8
, w
hich
sho
w
s
t
he
b
r
ea
k
s
iden
t
ified
in
[
19
]
du
r
ing
t
he
Gr
ea
t
Recession
.
L
o
g
D
a
ily
R
e
turns
1
Jan
2007
1
Jan
2008
1
Jan
2009
1
Jan
2010
1
Jan
2011
1
Jan
2012
-
0.1
-
0.05
0
0.05
0.1
0.45
Fi
g.
8
.
Re
g
ime
C
han
g
es
in
t
he
D
J
I
A
I
ndex
(indica
t
o
r
of
t
he
U
ni
t
ed
S
t
a
t
es
econom
y
)
iden
t
i
f
ied
b
y T
san
g
and
C
hen
[
19
]
.
P
r
e
d
i
ctab
ili
t
y
E
xt
e
rna
l
i
mpact
C
l
ass
i
ca
l
t
e
chn
i
qu
e
s
D
ata pr
e
pr
o
c
e
ss
i
ng
D
e
a
li
ng w
i
th chang
e
Charact
e
r
i
zat
i
o
n
o
f chang
e
P
r
e
d
i
ctab
ili
t
y
i
n pr
e
s
e
nc
e
o
f chang
e
R
e
g
i
m
e
charact
e
r
i
zat
i
o
n us
i
ng n
o
n sup
e
rv
i
s
e
d
l
e
arn
i
ng
T
rad
i
t
i
o
na
l
ML
D
ee
p L
e
arn
i
ng
E
ns
e
mb
l
e
s and m
e
ta
-l
e
arn
i
ng
E
v
o
l
v
i
ng
I
nt
e
lli
g
e
nt
Sy
st
e
ms
R
e
g
i
m
e
Chang
e
i
n
Fi
nanc
i
a
l
P
r
e
d
i
ct
i
o
n
Mach
i
n
e
l
e
arn
i
ng
R
e
g
i
m
e
Chang
e
B
as
i
c c
o
nc
e
pts
T
e
chn
i
qu
e
s
Fi
g.
7
. D
iscussion
on
t
he
r
esea
r
ch
pape
r
s
conside
r
ed
in
t
his
w
o
rk.
I
ntern
a
t
i
on
a
l
J
ourn
a
l
of
I
nter
ac
t
i
v
e
M
u
l
t
i
med
i
a a
nd
Art
i
f
i
c
i
a
l
I
nte
lli
g
en
c
e,
Vo
l
.
9
,
N
º
1
-
1
4
2 -
Regime
changes
challenge
in
v
es
t
o
r
s
,
ma
k
ing
t
hem
change
t
hei
r
t
r
ading
s
t
r
a
t
egies
as
t
he
collec
t
i
v
e
t
r
ading
beha
v
iou
r
of
t
he
ma
rk
e
t
changes
. D
iffe
r
en
t
examples
of
R
C
ha
v
e
been
co
v
e
r
ed
in
t
he
r
ecen
t
li
t
e
r
a
t
u
r
e
. D
a
v
ies
[
53
]
anal
y
sed
diffe
r
en
t
cases
and
consequences
of
r
egime
changes
in
t
he
Gr
ea
t
Recession
t
ha
t
impac
t
ed
se
v
e
r
al
asse
t
classes
such
as
equi
t
ies
,
bonds
,
commodi
t
ies
and
cu
rr
encies
a
t
mic
r
o
and
mac
r
oeconomic
le
v
els
. H
amil
t
on
[
63
]
obse
rv
ed
al
t
e
r
na
t
ing
pa
tt
e
r
ns
be
t
w
een
s
t
ead
y
and
t
u
r
bulen
t
pe
r
iods
since
t
he
Second
W
o
r
ld
W
a
r
and
subsequen
t
r
ecessions
b
y
loo
k
ing
a
t
U
S
unemplo
y
men
t
r
a
t
es
.
A
ng
and
T
imme
r
mann
[
15
]
iden
t
ified
c
y
clic
changes
in
t
he
beha
v
iou
r
of
asse
t
p
r
ices
and
mean
, v
ola
t
ili
t
y
and
co
rr
ela
t
ion
pa
tt
e
r
ns
in
s
t
oc
k
r
e
t
u
r
ns
du
r
ing
t
he
Gr
ea
t
Recession
and
t
he
1973
oil
c
r
isis
. Kr
i
t
zman
e
t
al
.
[
29
]
disco
v
e
r
ed
t
ha
t
in
v
es
t
o
r
s
could
benefi
t
f
r
om
ha
v
ing
diffe
r
en
t
asse
t
alloca
t
ion
s
t
r
a
t
egies
in
diffe
r
en
t
ma
rk
e
t
r
egimes
t
o
minimise
losses
.
M
an
y
o
t
he
r
s
t
udies
conside
r
t
hese
d
r
as
t
ic
changes
an
in
t
r
insic
cha
r
ac
t
e
r
is
t
ic
of
financial
da
t
a
t
ha
t
migh
t
be
caused
b
y
significan
t
e
v
en
t
s
,
and
t
hus
,
t
hese
w
ill
be
obse
rv
able
no
t
onl
y
in
p
r
ices
and
economic
v
a
r
iables
bu
t
also
in
o
t
he
r k
inds
of
public
info
r
ma
t
ion
[
52
]
,
[
58
]
,
[
59
]
,
[
62
]
,
[
63
]
,
[
92
]
. H
amil
t
on
[
58
]
p
r
oposed
a
t
ime
-
se
r
ies
based
app
r
oach
[
26
]
t
o
cap
t
u
r
e
nonlinea
r
effec
t
s
li
k
e
R
C,
iden
t
if
y
ma
rk
e
t
b
r
ea
k
s
and
hidden
changes
in
economic
c
y
cles
k
no
w
n
as
t
he
r
egime
-
s
w
i
t
ching
model
[
59
]
. T
his
model
,
also
k
no
w
n
as
t
he
M
a
rk
o
v-
s
w
i
t
ching
model
,
is
fi
tt
ed
t
o
obse
rv
a
t
ions
follo
w
ing
diffe
r
en
t
pa
tt
e
r
ns
in
diffe
r
en
t
pe
r
iods
and
is
mainl
y
applied
t
o
r
ecognise
lo
w
v
ola
t
ili
t
y r
egimes
w
i
t
h
economic
g
r
o
w
t
h
v
s
high
v
ola
t
ili
t
y
pe
r
iods
w
i
t
h
economic
con
t
r
ac
t
ions
[
116
]
.A
ng
and
T
imme
r
mann
[
15
]
applied
t
hese
models
t
o
p
r
edic
t
in
t
e
r
es
t
r
a
t
es
and
equi
t
y
and
fo
r
eign
exchange
r
e
t
u
r
ns
. T
he
y
discussed
ho
w
t
o
model
R
C
s
fo
r
t
hese
t
ime
se
r
ies
models
.
B
.
A
pproac
h
e
s
Ba
s
ed
on
Mac
hi
ne
L
earn
i
n
g
(Q
2
)
Tr
adi
t
ional
s
t
a
t
is
t
ical
me
t
hods
t
end
t
o
model
and
p
r
edic
t
fu
t
u
r
e
da
t
a
based
on
t
he
assump
t
ion
t
ha
t
t
he
t
ime
se
r
ies
unde
r
s
t
ud
y
is
gene
r
a
t
ed
f
r
om
a
linea
r
p
r
ocess
w
i
t
h
fea
t
u
r
es
no
r
mall
y
dis
t
r
ibu
t
ed
. T
his
p
r
esen
t
s
challenges
since
financial
da
t
a
is
cha
r
ac
t
e
r
ised
b
y
nonlinea
r
i
t
y
and
non
-
s
t
a
t
iona
r
i
t
y
besides
a
high
le
v
el
of
unce
r
t
ain
t
y
and
noise
[
82
]
.
A
diffe
r
en
t
app
r
oach
t
o
pe
r
fo
r
ming
financial
fo
r
ecas
t
s
is
t
he
use
of
ML
t
echniques
.
Se
v
e
r
al
li
t
e
r
a
t
u
r
e
r
e
v
ie
w
s
sho
w
t
he
benefi
t
s
of
t
hese
t
echniques
agains
t
t
r
adi
t
ional
me
t
hods
[
5
]
,
[
98
]
,
[
99
]
,
[
105
]
,
su
r
p
r
ising
p
r
ac
t
i
t
ione
r
s
b
y
con
t
r
adic
t
ing
ea
r
l
y
t
heo
r
ies
li
k
e
t
he
r
andom
w
al
k,
and
efficien
t
ma
rk
e
t
h
y
po
t
hesis
(E
MH
)
[
5
]
,
[
25
]
,
[
95
]
. M
achine
lea
r
ning
algo
r
i
t
hms
can
handle
nonlinea
r r
ela
t
ionships
w
i
t
hou
t
p
r
io
r k
no
w
ledge
[
7
]
,
[
144
]
,
ou
t
pe
r
fo
r
ming
t
r
adi
t
ional
t
ime
se
r
ies
me
t
hods
[
10
]
[
13
]
.
D
iffe
r
en
t
r
esea
r
ch
w
o
rk
s
f
r
om
t
he
economic
li
t
e
r
a
t
u
r
e
ha
v
e
ei
t
he
r
used
t
echnical
indica
t
o
r
s
o
r r
a
w
p
r
ices
and
r
e
t
u
r
ns
. T
echnical
indica
t
o
r
s
a
r
e
able
t
o
sho
w
beha
v
iou
r
al
pa
tt
e
r
ns
among
t
r
ade
r
s
and
t
hus
p
r
o
v
ide
an
ex
t
r
a
le
v
el
of
signal
t
o
p
r
edic
t
i
v
e
models
. T
hese
can
be
v
aluable
t
o
au
t
oma
t
e
t
he
beha
v
iou
r
of
sho
r
t
-
t
e
r
m
t
r
ade
r
s
[
5
]
.
I
n
an
y
case
,
mos
t
of
t
he
economics
li
t
e
r
a
t
u
r
e
has
focused
on
linea
r
p
r
ocesses
t
ha
t
ma
y
no
t
ha
v
e
been
able
t
o
ex
t
r
ac
t
r
ele
v
an
t
info
r
ma
t
ion
no
r
infe
r
complex
r
ela
t
ionships
among
t
echnical
indica
t
o
r
s
w
he
r
e
some
ne
w
ML
me
t
hods
could
[
108
]
.
Some
of
t
he
li
t
e
r
a
t
u
r
e
r
e
v
ie
w
s
al
r
ead
y
ci
t
ed
desc
r
ibe
common
t
echnical
indica
t
o
r
s
used
fo
r
s
t
oc
k
ma
rk
e
t
v
alue
and
t
r
end
p
r
edic
t
ion
. M
an
y
of
t
hese
pape
r
s
,
li
k
e
[
7
]
,
ha
v
e
sho
w
n
t
ha
t
diffe
r
en
t
p
r
e
-
p
r
ocessing
s
t
eps
,
li
k
e
t
he
f
r
equenc
y
le
v
el
of
t
he
inpu
t
da
t
a
,
can
impac
t
i
t
s
p
r
edic
t
abili
t
y. W
hile
a
common
app
r
oach
in
t
his
r
ega
r
d
is
da
t
a
no
r
malisa
t
ion
,
in
t
he
li
t
e
r
a
t
u
r
e
on
da
t
a
s
t
r
eam
mining
,
da
t
a
no
r
malisa
t
ion
is
no
t
a
usual
p
r
ac
t
ice
since
maximum
and
minimum
v
alues
fo
r
each
a
tt
r
ibu
t
e
in
t
he
da
t
a
s
t
r
eam
a
r
e
un
k
no
w
n
befo
r
ehand
[
134
]
.
A
u
t
ho
r
s
li
k
e
Pa
t
el
e
t
al
.
[
13
]
disc
r
e
t
ise
fea
t
u
r
es
based
on
t
he
human
app
r
oach
t
o
in
v
es
t
ing
and
de
r
i
v
ing
t
he
t
echnical
indica
t
o
r
s
using
assump
t
ions
f
r
om
t
he
s
t
oc
k
ma
rk
e
t
. T
his
la
tt
e
r
app
r
oach
,
t
hough
,
in
t
r
oduces
human
bias
in
t
he
p
r
ocess
. T
his
is
t
he
opposi
t
e
w
a
y
t
o
app
r
oach
t
he
p
r
oblem
of
t
r
end
p
r
edic
t
ion
if
w
e
compa
r
e
i
t
t
o
r
ecen
t
deep
lea
r
ning
s
t
r
a
t
egies
t
ha
t
feed
dozens
of
au
t
oma
t
icall
y
gene
r
a
t
ed
indica
t
o
r
s
[
109
]
.
O
v
e
r
all
,
t
he
abo
v
e
-
men
t
ioned
li
t
e
r
a
t
u
r
e
r
e
v
ie
w
s
confi
r
m
t
ha
t
ML
t
echniques
can
be
used
t
o
p
r
edic
t
p
r
ice
changes
,
bu
t
t
his
en
t
i
r
el
y
depends
on
t
he
t
ime
ho
r
izon
and
efficienc
y
of
t
he
ma
rk
e
t
in
t
he
pe
r
iod
p
r
edic
t
ed
. C
a
v
alcan
t
e
e
t
al
.
[
3
]
p
r
o
v
ided
ano
t
he
r
in
t
e
r
es
t
ing
r
e
v
ie
w
of
p
r
e
-
p
r
ocessing
and
clus
t
e
r
ing
t
echniques
used
in
t
he
financial
domain
t
o
fo
r
ecas
t
fu
t
u
r
e
ma
rk
e
t
mo
v
emen
t
s
. T
he
y
highligh
t
ed
t
he
r
ele
v
ance
of
concep
t
d
r
if
t
s
in
financial
ma
rk
e
t
s
and
sugges
t
ed
t
ha
t
t
he
da
t
a
s
t
r
eam
mining
li
t
e
r
a
t
u
r
e
is
of
g
r
ea
t
impo
r
t
ance
in
fu
t
u
r
e
r
esea
r
ch
due
t
o
t
he
non
-
s
t
a
t
iona
r
i
t
y
and
e
v
olu
t
ion
of
financial
ma
rk
e
t
s
[
91
]
.
I
n
compu
t
a
t
ional
finance
,
changes
in
t
he
beha
v
iou
r
of
t
he
ma
rk
e
t
a
r
e
no
r
mall
y r
efe
rr
ed
t
o
as
r
egime
changes
o
r
s
w
i
t
ches
[
15
]
,
[
61
]
,
[
67
]
,
s
t
r
uc
t
u
r
al
b
r
ea
k
s
o
r
changes
[
51
]
,
[
54
]
,
[
56
]
,
[
56
]
, v
ola
t
ili
t
y
shif
t
s
[
50
]
,
s
w
i
t
ching
p
r
ocesses
[
60
]
o
r
ma
rk
e
t
s
t
a
t
es
[
18
]
.
I
n
t
his
k
ind
of
da
t
a
,
long
pe
r
iods
of
s
t
abili
t
y
migh
t
be
in
t
e
rr
up
t
ed
b
y
sho
r
t
episodes
of
ab
r
up
t
changes
[
61
]
.
T
hese
changes
ma
y
o
r
no
t
be
t
r
ansi
t
o
ry
since
a
ne
w
l
y
adop
t
ed
beha
v
iou
r
in
p
r
ice
d
y
namics
, r
eflec
t
ed
as
pa
r
t
of
t
he
mean
r
e
t
u
r
ns
,
t
hei
r v
ola
t
ili
t
y
o
r
co
rr
ela
t
ion
among
t
hem
ma
y
pe
r
sis
t
fo
r
se
v
e
r
al
pe
r
iods
. T
imel
y r
ecogni
t
ion
of
t
hese
sudden
beha
v
iou
r
al
changes
in
ma
rk
e
t
s
can
significan
t
l
y
lo
w
e
r
t
he
r
is
k
of
financial
exposu
r
e
. T
his
has
inspi
r
ed
t
he
ma
t
e
r
ialisa
t
ion
of
t
echniques
such
as
r
egime
-
s
w
i
t
ching
models
in
t
he
financial
li
t
e
r
a
t
u
r
e
, w
hich
w
o
rk
unde
r
t
he
p
r
emise
t
ha
t
ne
w
d
y
namics
of
p
r
ice
r
e
t
u
r
ns
and
fundamen
t
als
pe
r
sis
t
fo
r
se
v
e
r
al
pe
r
iods
af
t
e
r
a
change
. A k
e
y
elemen
t
in
t
hese
models
is
iden
t
if
y
ing
w
he
t
he
r
t
he
exac
t
ma
rk
e
t
r
egimes
r
eoccu
r
o
v
e
r
t
ime
(e
.
g
.
ac
r
oss
r
ecessions
o
r
pe
r
iods
of
economic
g
r
o
w
t
h)
o
r
if
ne
w r
egimes
de
v
ia
t
e
o
r
ha
v
e
e
v
ol
v
ed
f
r
om
p
r
e
v
ious
ones
[
15
]
.
T
he
p
r
edic
t
ion
of
fu
t
u
r
e
v
alues
in
financial
ma
rk
e
t
s
and
t
he
de
t
ec
t
ion
of
r
egime
changes
in
da
t
a
s
t
r
eams
w
i
t
h
t
empo
r
al
dependence
a
r
e
common
applica
t
ion
a
r
eas
fo
r
s
t
a
t
is
t
ical
me
t
hods
and
ML.
P
r
e
v
ious
r
esea
r
ch
r
epo
r
t
s
high
accu
r
ac
y
in
fo
r
ecas
t
ing
p
r
ice
changes
w
i
t
h
ad
v
anced
t
echniques
and
t
he
feasibili
t
y
of
ma
k
ing
p
r
ofi
t
s
using
t
hese
p
r
edic
t
ions
agains
t
t
he
E
MH, w
hich
poin
t
s
t
o
unbea
t
able
ma
rk
e
t
s
.
A
n
al
t
e
r
na
t
i
v
e
t
heo
ry
is
t
he
adap
t
i
v
e
ma
rk
e
t
h
y
po
t
hesis
(
AMH
)
[
57
]
,
in
t
r
oduced
b
y A
nd
r
e
w L
o
in
2004
. T
his
t
heo
ry, w
i
t
h
empi
r
ical
e
v
idence
in
a
inc
r
easing
numbe
r
of
r
esea
r
ch
w
o
rk
s
[
68
]
,
combines
t
he
E
MH w
i
t
h
p
r
inciples
f
r
om
beha
v
iou
r
al
finance
,
allo
w
ing
t
he
ideas
of
ma
rk
e
t
efficienc
y
and
inefficiencies
t
o
co
-
exis
t
. U
nde
r
t
he
AMH,
t
he
efficienc
y
of
a
ma
rk
e
t
e
v
ol
v
es
as
ma
rk
e
t
pa
r
t
icipan
t
s
adap
t
t
o
an
en
v
i
r
onmen
t
t
ha
t
changes
con
t
inuousl
y.
I
n
t
his
r
ega
r
d
,
pa
r
t
icipan
t
s
r
el
y
on
heu
r
is
t
ics
t
o
ma
k
e
t
hei
r
in
v
es
t
men
t
choice
,
leading
t
o
mos
t
l
y
r
a
t
ional
ma
rk
e
t
s
unde
r
t
hose
heu
r
is
t
ics
(li
k
e
t
he
E
MH
)
. T
he
main
diffe
r
ence
is
a
t
t
he
t
ime
of
majo
r
beha
v
iou
r
al
shif
t
s
in
t
he
ma
rk
e
t
pa
r
t
icipan
t
s
,
as
in
economic
shoc
k
s
o
r
c
r
ises
.
I
n
t
his
case
,
t
he
AMH
conside
r
s
a
ma
rk
e
t
t
ha
t
e
v
ol
v
es
,
and
t
he
ini
t
iall
y
adap
t
i
v
e
heu
r
is
t
ics
ma
y
become
s
t
a
t
ic
in
ce
r
t
ain
ma
rk
e
t
si
t
ua
t
ions
. C
onsequen
t
l
y,
t
he
E
MH
ma
y
no
t
con
t
inue
unde
r
pe
r
iods
of
abno
r
mal
condi
t
ions
,
s
t
r
ess
o
r
ab
r
up
t
changes
in
t
he
ma
rk
e
t
. H
ence
,
financial
ma
rk
e
t
s
ma
y
be
p
r
edic
t
able
in
specific
pe
r
iods
,
as
discussed
b
y L
o
[
49
]
. T
he
r
efo
r
e
,
con
v
e
r
gence
t
o
ma
rk
e
t
efficienc
y
is
nei
t
he
r
gua
r
an
t
eed
no
r
li
k
el
y
t
o
occu
r. T
he
le
v
el
of
efficienc
y
depends
on
t
he
ma
rk
e
t
pa
r
t
icipan
t
s
and
t
he
ma
rk
e
t
condi
t
ions
a
t
t
ha
t
t
ime
.
One
of
t
he
fe
w
financial
s
t
udies
ci
t
ing
concep
t
d
r
if
t
explici
t
l
y
can
be
found
in
a
r
ecen
t
w
o
rk
au
t
ho
r
ed
b
y M
asegosa
e
t
al
.
[
36
]
. T
he
y
anal
y
sed
da
t
a
f
r
om
t
he
Gr
ea
t
Recession
and
claimed
t
ha
t
economic
R
e
g
u
l
a
r
Iss
ue
-
1
4
3 -
changes
du
r
ing
t
his
pe
r
iod
manifes
t
ed
as
concep
t
d
r
if
t
s
in
t
hei
r
gene
r
a
t
i
v
e
p
r
ocesses
. A
n
in
t
e
r
media
t
e
example
of
t
ry
ing
t
o
p
r
edic
t
financial
c
r
ises
using
ML
me
t
hods
can
be
found
in
[
55
]
, w
he
r
e
t
he
au
t
ho
r
s
s
t
udied
possible
con
t
agion
r
is
k
s
be
t
w
een
financial
ma
rk
e
t
s
t
ha
t
could
t
r
igge
r
financial
c
r
ises
t
o
signal
w
a
r
nings
a
t
an
ea
r
l
y
s
t
age
.
M
o
r
e
r
ecen
t
l
y,
Yang
e
t
al
.
[
137
]
anal
y
sed
t
he
impac
t
of
concep
t
d
r
if
t
in
business
p
r
ocesses
. M
o
r
e
specificall
y,
t
he
y
modelled
t
he
r
esponse
t
o
concep
t
d
r
if
t
as
a
sequen
t
ial
decision
-
ma
k
ing
p
r
oblem
b
y
combing
a
hie
r
a
r
chical
M
a
rk
o
v
model
and
a
M
a
rk
o
v
decision
p
r
ocess
. M
a
r
t
ín
e
t
al
.
[
145
]
also
deal
t
w
i
t
h
s
t
r
uc
t
u
r
al
changes
in
t
r
oducing
a
t
r
ading
s
y
s
t
em
based
on
g
r
amma
t
ical
e
v
olu
t
ion
t
ha
t
commu
t
es
be
t
w
een
an
ac
t
i
v
e
model
and
a
candida
t
e
one
t
o
inc
r
ease
pe
r
fo
r
mance
.
O
t
he
r
t
w
o
k
e
y r
esea
r
ch
pieces
in
t
his
r
ega
r
d
a
r
e
t
he
w
o
rk
s
b
y
T
sang
and
C
hen
[
19
]
and
M
"unnix
e
t
al
.
[
18
]
, w
hich
p
r
oposed
mechanisms
t
o
iden
t
if
y
poin
t
s
of
d
r
as
t
ic
changes
in
financial
t
ime
se
r
ies
. T
he
fo
r
me
r
used
s
t
a
t
is
t
ical
-
based
and
t
r
adi
t
ional
ML
(e
.
g
.
nai
v
e
Ba
y
es)
app
r
oaches
t
o
classif
y
no
r
mal
v
e
r
sus
abno
r
mal
r
egimes
. T
he
y
p
r
oposed
a
f
r
ame
w
o
rk
based
on
t
he
change
speed
of
p
r
ice
r
e
t
u
r
ns
and
t
he
deg
r
ee
of
changes
t
o
v
isualise
and
disc
r
imina
t
e
be
t
w
een
diffe
r
en
t
ma
rk
e
t
r
egimes
depending
on
t
he
v
ola
t
ili
t
y
of
t
hei
r
p
r
ice
r
e
t
u
r
ns
. T
he
la
tt
e
r
[
18
]
v
isualised
diffe
r
ences
in
t
he
co
rr
ela
t
ion
s
t
r
uc
t
u
r
e
of
t
he
p
r
ice
r
e
t
u
r
ns
ac
r
oss
asse
t
s
in
t
he
S
&
P500
du
r
ing
t
he
Gr
ea
t
Recession
.
T
he
y
ex
t
ended
t
he
selec
t
ion
t
o
a
sample
f
r
om
1992
t
o
2010
,
iden
t
if
y
ing
eigh
t
ma
rk
e
t
s
t
a
t
es
r
epea
t
ing
beha
v
iou
r
al
changes
o
v
e
r
t
ime
.
Se
v
e
r
al
app
r
oaches
f
r
om
t
he
deep
lea
r
ning
(
DL
)
field
(e
.
g
.
RNNs)
ha
v
e
also
t
r
ied
t
o
face
t
he
p
r
oblem
of
concep
t
changes
w
hen
lea
r
ning
con
t
inuousl
y.
[
113
]
,
[
126
]
,
[
128
]
,
[
130
]
. T
hese
ad
v
ances
ha
v
e
been
successfull
y
t
r
ansfe
rr
ed
t
o
t
he
financial
domain
,
as
discussed
in
r
ecen
t
su
rv
e
y
s
b
y
Ozba
y
oglu
e
t
al
.
[
9
]
,
and
L
i
and
Bas
t
os
[
109
]
.
M
os
t
of
t
hese
men
t
ioned
r
esea
r
ch
w
o
rk
s
focus
on
t
he
li
k
elihood
of
dail
y
o
r
mon
t
hl
y
changes
, w
he
r
e
r
e
t
r
aining
a
model
is
a
feasible
t
as
k. A
s
of
t
oda
y,
t
he
amoun
t
of
r
esea
r
ch
de
v
o
t
ed
t
o
seasonali
t
y
and
changes
a
t
t
he
in
t
r
ada
y
le
v
el
is
significan
t
l
y
mo
r
e
limi
t
ed
[
75
]
,
[
89
]
,
[
101
]
,
[
103
]
,
[
107
]
. T
he
compu
t
a
t
ional
cos
t
of
ML
and
s
t
a
t
is
t
ical
me
t
hods
,
t
oge
t
he
r w
i
t
h
t
he
inhe
r
en
t
highe
r
complexi
t
y
de
r
i
v
ed
f
r
om
t
he
need
t
o
manage
la
r
ge
amoun
t
s
of
da
t
a
a
t
t
hese
r
esolu
t
ions
,
ma
k
es
k
eeping
models
up
t
o
da
t
e
mo
r
e
challenging
.
W
hile
i
t
s
applica
t
ion
t
o
high
-
f
r
equenc
y
ma
rk
e
t
s
is
s
t
ill
an
open
p
r
oblem
, r
ecen
t
r
esea
r
ch
w
o
rk
s
a
r
e
ma
k
ing
p
r
og
r
ess
in
unde
r
s
t
anding
ho
w
t
o
appl
y ML
t
o
in
t
r
ada
y r
esolu
t
ions
. A
mong
t
hem
, w
e
could
men
t
ion
t
he
one
p
r
esen
t
ed
b
y
Si
r
ignano
and
C
on
t
[
73
]
, w
ho
claimed
t
ha
t
financial
da
t
a
a
t
high
f
r
equencies
exhibi
t
s
t
y
lised
fac
t
s
and
ma
y
hold
lea
r
nable
s
t
a
t
iona
ry
pa
tt
e
r
ns
o
v
e
r
long
pe
r
iods
. A
no
t
he
r r
ele
v
an
t
s
t
ud
y
is
t
he
one
au
t
ho
r
ed
b
y
Shin
t
a
t
e
and
Pichl
[
76
]
, w
ho
r
e
v
ie
w
ed
mode
r
n
ML
and
DL
app
r
oaches
applied
t
o
high
-
f
r
equenc
y
t
r
ading
a
t
t
he
minu
t
e
le
v
el
.
Recen
t
l
y,
se
v
e
r
al
r
esea
r
ch
w
o
rk
s
ha
v
e
app
r
oached
t
he
p
r
oblem
of
t
ime
-
changing
beha
v
iou
r
s
using
non
-
supe
rv
ised
ML
me
t
hods
[
127
]
,
[
130
]
,
[
135
]
.
I
n
t
hese
,
mic
r
o
-
clus
t
e
r
s
o
r
la
t
en
t
fea
t
u
r
es
ma
y
be
used
t
o
r
ep
r
esen
t
a
summa
ry
of
t
he
incoming
da
t
a
and
r
educe
t
he
compu
t
a
t
ional
cos
t
s
of
co
rr
ela
t
ing
full
da
t
a
dis
t
r
ibu
t
ions
. A
manne
r
of
doing
t
his
is
using
model
-
based
clus
t
e
r
ing
app
r
oaches
.
T
hese
algo
r
i
t
hms
find
models
t
ha
t
fi
t
inpu
t
da
t
a
and
a
r
e
also
r
obus
t
t
o
t
he
p
r
esence
of
noise
[
146
]
,
[
147
]
.
Fo
r
ins
t
ance
,
expec
t
a
t
ion
maximisa
t
ion
(E
M
)
[
148
]
fi
t
s
a
mix
t
u
r
e
of
G
aussian
dis
t
r
ibu
t
ions
t
o
t
he
da
t
a
[
133
]
. C
hiu
and
M
in
k
u
[
141
]
used
i
t
in
concep
t
d
r
if
t
handling
based
on
clus
t
e
r
ing
in
t
he
model
space
(
CDCM
S)
t
o
c
r
ea
t
e
concep
t
r
ep
r
esen
t
a
t
ions
and
k
eep
a
di
v
e
r
se
ensemble
lea
r
ne
r. Z
heng
e
t
al
.
[
123
]
r
elied
on
i
t
t
o
minimise
in
t
r
a
-
clus
t
e
r
dispe
r
sion
and
clus
t
e
r
impu
r
i
t
y.
T
sang
and
C
hen
[
19
]
applied
t
he
Baum–
W
elch
algo
r
i
t
hm
,
a
special
case
of
E
M,
t
o
bo
t
h
de
t
ec
t
t
he
t
ime
of
a
change
poin
t
and
p
r
edic
t
t
he
nex
t
s
t
a
t
e
(o
r
concep
t
)
of
financial
da
t
a
using
a
hidden
M
a
rk
o
v
model
(
HMM
)
. G
omes
e
t
al
.
[
122
]
also
h
y
po
t
hesised
abou
t
using
Baum–
W
elch
in
conjunc
t
ion
w
i
t
h
HMM
s
fo
r
con
t
inuous
lea
r
ning
p
r
oblems
.
Baum
W
elch
has
been
used
in
t
he
financial
domain
t
oge
t
he
r w
i
t
h
o
t
he
r
specific
v
e
r
sions
of
E
M
and
G
aussian
mix
t
u
r
e
models
(
GMM
)
t
o
fo
r
ecas
t
change
di
r
ec
t
ion
in
s
t
oc
k
p
r
ices
[
72
]
,
[
87
]
and
t
o
r
ep
r
esen
t
ma
rk
e
t
r
egimes
[
19
]
,
[
29
]
,
[
64
]
,
[
86
]
.
A
se
t
of
r
ele
v
an
t
t
echniques
f
r
om
t
he
ML
li
t
e
r
a
t
u
r
e
in
t
his
r
ega
r
d
a
r
e
p
r
o
t
o
t
y
pe
gene
r
a
t
ion
t
echniques
such
as
lea
r
ning
v
ec
t
o
r
quan
t
isa
t
ion
[
65
]
,
[
96
]
,
[
102
]
, w
hich
ha
v
e
been
p
r
o
v
en
t
o
be
useful
fo
r
da
t
a
pa
r
t
i
t
ioning
and
model
selec
t
ion
in
t
he
financial
domain
. C
houdhu
ry
e
t
al
.
[
71
]
,
and
Pa
v
lidis
e
t
al
.
[
69
]
use
a
combina
t
ion
of
clus
t
e
r
ing
and
fo
r
ecas
t
ing
algo
r
i
t
hms
t
o
model
t
he
dis
t
r
ibu
t
ion
of
financial
da
t
a
.
Rega
r
ding
t
he
fi
r
s
t
s
t
ep
,
t
he
fo
r
me
r
au
t
ho
r
s
use
a
t
w
o
-
la
y
e
r
abs
t
r
ac
t
ion
t
ha
t
clus
t
e
r
s
s
t
oc
k
s
using
self
-
o
r
ganising
maps
(SO
M
)
t
o
t
hen
r
el
y
on
K-
means
t
o
ob
t
ain
clus
t
e
r
s
of
p
r
o
t
o
t
y
pes
. T
he
la
tt
e
r
conside
r
s
t
h
r
ee
diffe
r
en
t
unsupe
rv
ised
algo
r
i
t
hms
t
o
iden
t
if
y
ma
rk
e
t
s
t
a
t
es
:
g
r
o
w
ing
neu
r
al
gas
(
G
N
G
)
,
densi
t
y-
based
spa
t
ial
clus
t
e
r
ing
of
applica
t
ions
w
i
t
h
noise
(
D
BS
CA
N)
,
and
U
nsupe
rv
ised
k-W
indo
w
s
.
Once
t
he
ma
rk
e
t
s
t
a
t
es
a
r
e
iden
t
ified
,
t
he
y
use
feed
-
fo
rw
a
r
d
neu
r
al
ne
t
w
o
rk
s
t
o
ma
k
e
p
r
edic
t
ions
.
I
n
t
he
las
t
y
ea
r
s
,
se
v
e
r
al
online
inc
r
emen
t
al
algo
r
i
t
hms
ha
v
e
used
t
hese
t
echniques
t
o
adap
t
dis
t
inc
t
lea
r
ne
r
s
t
o
diffe
r
en
t
c
y
cles
o
r
seasonal
beha
v
iou
r
s
in
a
da
t
a
s
t
r
eam
.
I
n
t
his
r
ega
r
d
,
t
he
use
of
online
ensembles
,
using
non
-
supe
rv
ised
lea
r
ning
t
o
r
ep
r
esen
t
diffe
r
en
t
beha
v
iou
r
al
pa
tt
e
r
ns
[
135
]
,
[
141
]
o
r
supe
rv
ised
lea
r
ning
t
o
t
r
ain
a
pool
of
classifie
r
s
w
i
t
h
high
p
r
edic
t
i
v
e
accu
r
ac
y
unde
r
diffe
r
en
t
condi
t
ions
[
34
]
,
[
117
]
,
ha
v
e
ob
t
ained
s
t
a
t
e
-
of
-
t
he
-
a
r
t
r
esul
t
s
adap
t
ing
t
o
diffe
r
en
t
s
t
a
t
es
in
t
he
beha
v
iou
r
of
da
t
a
s
t
r
eams
in
man
y
domains
,
including
finance
[
17
]
.
M
e
t
a
lea
r
ne
r
s
o
v
e
r
da
t
a
s
t
r
eams
a
r
e
a
r
ela
t
ed
subfield
of
ML
of
inc
r
easing
popula
r
i
t
y w
he
r
e
a
pool
of
fo
r
me
r
classifie
r
s
is
managed
and
r
eused
w
hen
t
he
s
t
a
t
e
o
r
concep
t
of
t
he
s
t
r
eam
changes
. T
his
subfield
is
inspi
r
ed
b
y
t
he
human
lea
r
ning
s
y
s
t
em
t
ha
t
r
euses
p
r
e
v
ious
k
no
w
ledge
t
o
lea
r
n
ne
w
t
as
k
s
,
no
t
s
t
a
r
t
ing
f
r
om
ze
r
o
e
v
e
ry
t
ime
. A
l
t
hough
me
t
a
-
lea
r
ne
r
s
ha
v
e
no
t
been
w
idel
y
applied
t
o
t
he
financial
domain
y
e
t
,
t
hei
r
logic
r
esembles
app
r
oaches
applied
t
o
finance
as
E
M
o
r
Baum
-W
elch
,
bu
t
fo
r
con
t
inuous
lea
r
ning
domains
w
he
r
e
models
a
r
e
al
w
a
y
s
up
-
t
o
-
da
t
e
and
t
hus
beha
v
e
smoo
t
hl
y
in
case
of
s
t
r
uc
t
u
r
al
b
r
ea
k
s
.
Fo
r
ins
t
ance
, A
bad
e
t
al
.
[
120
]
p
r
oposed
a
me
t
a
-
lea
r
ne
r
t
ha
t
used
hidden
M
a
rk
o
v
models
(
HMM
)
t
o
p
r
edic
t
t
he
sequence
of
change
be
t
w
een
disc
r
e
t
e
concep
t
s
. T
hei
r
app
r
oach
used
fuzz
y
logic
r
ules
t
o
compa
r
e
classifie
r
s
t
o
r
euse
fo
r
me
r
models
. M
aslo
v
e
t
al
.
[
139
]
p
r
oposed
a
me
t
hod
t
o
use
pa
tt
e
r
ns
acqui
r
ed
du
r
ing
p
r
e
v
ious
changes
and
assumed
a
G
aussian
dis
t
r
ibu
t
ion
fo
r
t
he
du
r
a
t
ion
of
t
he
changes
t
o
p
r
edic
t
t
he
t
ime
of
t
he
nex
t
change
poin
t
.
C
a
r
t
a
e
t
al
.
in
[
83
]
r
ecen
t
l
y
combined
t
he
use
of
me
t
a
lea
r
ne
r
s
w
i
t
h
deep
r
einfo
r
cemen
t
lea
r
ning
t
o
p
r
oduce
t
r
ading
s
t
r
a
t
egies
and
maximise
p
r
ofi
t
s
ope
r
a
t
ing
in
S
t
anda
r
d
&
Poo
r
’s
500
fu
t
u
r
e
ma
rk
e
t
s
and
t
he
J
.
P
. M
o
r
gan
and
M
ic
r
osof
t
s
t
oc
k
s
.
M
e
t
a
-
lea
r
ning
app
r
oaches
ha
v
e
inbuil
t
s
t
r
a
t
egies
t
o
decide
on
w
hen
t
o
t
r
ain
, w
ha
t
model
t
o
r
eplace
and
w
hen
, w
hen
t
o
fo
r
ge
t
(p
r
une)
a
lea
r
ne
r
and
w
hen
t
o
c
r
ea
t
e
one
[
114
]
,
[
136
]
b
y
using
t
he
e
v
alua
t
ion
pe
r
fo
r
mance
me
t
r
ics
of
ac
t
i
v
e
and
fo
r
me
r
models
[
149
]
. T
hese
,
t
hus
,
a
r
e
a
closel
y r
ela
t
ed
r
esea
r
ch
a
r
ea
t
o
e
v
ol
v
ing
in
t
elligen
t
s
y
s
t
ems
(E
I
S)
[
118
]
,
[
129
]
,
[
150
]
and
e
v
ol
v
ing
fuzz
y
s
y
s
t
ems
[
125
]
[
132
]
[
143
]
[
119
]
.
E
I
S
, w
hich
a
r
e
also
online
and
inc
r
emen
t
al
s
y
s
t
ems
,
can
adap
t
t
hemsel
v
es
t
o
concep
t
d
r
if
t
s
of
diffe
r
en
t
na
t
u
r
es
on
-
t
he
-
fl
y
t
h
r
ough
adap
t
i
v
e
fuzz
y r
ules
[
140
]
.
E
I
S
ha
v
e
al
r
ead
y
demons
t
r
a
t
ed
t
hei
r
abili
t
y
t
o
sol
v
e
diffe
r
en
t
k
inds
of
p
r
oblems
in
v
a
r
ious
applica
t
ion
domains
li
k
e
finance
[
90
]
,
[
97
]
,
[
151
]
.
T
hese
ha
v
e
achie
v
ed
g
r
ea
t
r
esul
t
s
in
classif
y
ing
non
-
s
t
a
t
iona
ry
t
ime
se
r
ies
[
24
]
,
[
37
]
,
[
38
]
.
I
ntern
a
t
i
on
a
l
J
ourn
a
l
of
I
nter
ac
t
i
v
e
M
u
l
t
i
med
i
a a
nd
Art
i
f
i
c
i
a
l
I
nte
lli
g
en
c
e,
Vo
l
.
9
,
N
º
1
-
1
44
-
Recen
t
E
I
S
app
r
oaches
can
w
o
rk
as
ensembles
of
r
ules
[
116
]
and
appl
y
me
t
a
-
cogni
t
i
v
e
scaffolding
t
heo
ry
fo
r
t
uning
t
he
lea
r
ned
model
inc
r
emen
t
all
y
in
w
ha
t
-
t
o
-
lea
r
n
, w
hen
-
t
o
-
lea
r
n
,
and
ho
w-
t
o
-
lea
r
n
[
138
]
.
I
n
fac
t
,
ensembles
a
r
e
k
no
w
n
fo
r
t
hei
r
good
r
esul
t
s
in
p
r
edic
t
ing
bo
t
h
c
y
clic
and
non
-
s
t
a
t
iona
ry
da
t
a
such
as
s
t
oc
k
p
r
ices
[
10
]
,
[
13
]
,
[
110
]
. T
hese
ha
v
e
also
in
t
r
oduced
t
he
abili
t
y
t
o
deal
w
i
t
h
r
ecu
rr
en
t
concep
t
s
explici
t
l
y
and
ha
v
e
bea
t
en
o
t
he
r
me
t
hods
a
t
p
r
edic
t
ing
t
he
S
&
P500
[
37
]
,
[
38
]
,
[
111
]
.
Fo
r
ins
t
ance
,
P
r
a
t
ama
e
t
al
.
emplo
y
ed
an
e
v
ol
v
ing
t
y
pe
-
2
r
ecu
rr
en
t
fuzz
y
neu
r
al
ne
t
w
o
rk
t
o
lea
r
n
inc
r
emen
t
all
y
and
handle
r
ecu
rr
ing
d
r
if
t
s
in
bo
t
h
[
124
]
and
[
38
]
.
I
n
an
y
case
,
t
he
r
e
is
s
t
ill
a
significan
t
gap
be
t
w
een
E
I
S
and
t
he
r
es
t
of
t
he
li
t
e
r
a
t
u
r
e
fo
r
da
t
a
s
t
r
eam
classifica
t
ion
.
T
his
line
of
w
o
rk
based
on
E
I
S
is
being
complemen
t
ed
b
y
o
t
he
r
s
t
udies
t
ha
t
combine
ensembles
and
o
t
he
r
e
v
olu
t
iona
ry
algo
r
i
t
hms
t
o
t
ac
k
le
concep
t
d
r
if
t
in
financial
applica
t
ions
li
k
e
[
84
]
. T
hese
au
t
ho
r
s
used
ensembles
of
t
r
ading
r
ules
e
v
ol
v
ed
using
g
r
amma
t
ical
e
v
olu
t
ion
t
o
manage
s
t
r
uc
t
u
r
al
change
in
t
he
S
t
anda
r
d
&
Poo
r
’s
500
index
.
V. Conclusions
a
nd
F
utur
e
Wor
k
T
he
applica
t
ion
of
A
I
t
o
compu
t
a
t
ional
finance
has
been
a
v
e
ry
ac
t
i
v
e
field
of
r
esea
r
ch
fo
r
decades
. A
mong
t
he
k
e
y
difficul
t
ies
iden
t
ified
in
t
he
li
t
e
r
a
t
u
r
e
on
financial
p
r
edic
t
ion
, w
e
can
men
t
ion
s
t
r
uc
t
u
r
al
change
. T
he
p
r
ice
gene
r
a
t
ion
p
r
ocess
of
financial
t
ime
se
r
ies
is
of
t
en
affec
t
ed
b
y
changes
of
diffe
r
en
t
na
t
u
r
es
.
Some
of
t
hese
changes
can
r
eoccu
r
o
v
e
r
t
ime
as
seasonal
pa
tt
e
r
ns
, w
hile
o
t
he
r
s
do
no
t
r
epea
t
,
being
ab
r
up
t
b
r
ea
k
s
in
t
he
non
-
s
t
a
t
iona
ry
p
r
ice
d
y
namics
.
T
his
s
t
ud
y
p
r
esen
t
ed
a
s
y
s
t
ema
t
ic
li
t
e
r
a
t
u
r
e
r
e
v
ie
w
of
machine
lea
r
ning
t
echniques
fo
r
financial
p
r
edic
t
ion
unde
r r
egime
changes
. A
v
a
r
ie
t
y
of
sou
r
ces
w
e
r
e
inspec
t
ed
t
o
pe
r
fo
r
m
an
exhaus
t
i
v
e
sea
r
ch
.
T
his
r
e
v
ie
w
included
:
Science
D
i
r
ec
t
,
I
EEE
D
igi
t
al
L
ib
r
a
ry, ACM
D
igi
t
al
L
ib
r
a
ry, T
a
y
lo
r
&
F
r
ancis
, W
ile
y
and
Sp
r
inge
rL
in
k.
Ou
t
of
a
t
o
t
al
of
140
r
ele
v
an
t
s
t
udies
,
t
hese
a
r
e
dis
t
r
ibu
t
ed
as
follo
w
s
:
i)
concep
t
d
r
if
t
o
r
online
lea
r
ning
r
ela
t
ed
(32
.
1
%
)
;
ii)
r
ela
t
ed
t
o
financial
li
t
e
r
ac
y
fo
r r
egime
changes
(15
.
7
%
)
,
and
iii)
ML
t
echniques
applied
t
o
s
t
oc
k
fo
r
ecas
t
ing
(52
.
1
%
)
.
T
he
r
esul
t
s
of
r
e
v
ie
w
ed
publica
t
ions
sho
w
t
ha
t
ML
has
p
r
o
v
en
t
o
be
a
po
w
e
r
ful
t
ool
t
o
t
ac
k
le
t
he
p
r
oblem
of
financial
p
r
edic
t
ion
unde
r
concep
t
d
r
if
t
, w
hich
w
e
define
as
s
t
r
uc
t
u
r
al
b
r
ea
k
s
,
t
ha
t
can
occu
r
a
t
an
y
f
r
equenc
y
le
v
el
. T
his
includes
solu
t
ions
based
on
diffe
r
en
t
algo
r
i
t
hms
t
ha
t
adap
t
t
hei
r
p
r
edic
t
ion
models
t
o
ne
w
ci
r
cums
t
ances
ei
t
he
r
t
h
r
ough
ne
w
model
gene
r
a
t
ion
o
r
managing
an
a
r
chi
v
e
of
fo
r
me
r
successful
models
.
I
n
t
his
r
ega
r
d
,
man
y
me
t
a
-
lea
r
ning
app
r
oaches
in
t
he
ML
li
t
e
r
a
t
u
r
e
r
el
y
on
non
-
supe
rv
ised
algo
r
i
t
hms
t
o
t
ry
t
o
iden
t
if
y
t
he
r
ecu
rr
ence
of
a
concep
t
and
r
e
t
r
ie
v
e
p
r
e
v
ious
models
o
r
de
t
ec
t
d
r
if
t
s
.
I
n
t
his
con
t
ex
t
,
t
he
use
of
sequen
t
ial
DL
models
such
as
RNNs
can
be
insufficien
t
t
o
t
ac
k
le
ab
r
up
t
changes
since
t
he
p
r
e
v
ious
ma
rk
e
t
d
y
namics
a
r
e
s
t
ill
in
memo
ry
in
t
he
models
impac
t
ing
p
r
edic
t
i
v
e
accu
r
ac
y.
I
n
con
t
r
as
t
,
t
he
model
lea
r
ns
t
he
ne
w r
egime
[
113
]
,
[
115
]
,
[
121
]
.
D
epending
on
t
he
f
r
equenc
y
in
v
ol
v
ed
, r
esea
r
che
r
s
sugges
t
ei
t
he
r
solu
t
ions
based
on
model
r
e
t
r
aining
ei
t
he
r
a
t
r
egula
r
in
t
e
rv
als
o
r
upon
de
t
ec
t
ion
of
changes
o
r
d
r
if
t
s
o
r
online
inc
r
emen
t
al
algo
r
i
t
hms
.
T
his
en
t
ails
ha
v
ing
up
-
t
o
-
da
t
e
models
w
i
t
h
t
he
use
of
fo
r
ge
tt
ing
mechanisms
t
o
a
v
oid
o
v
e
r
fi
tt
ing
and
adap
t
ing
t
o
ne
w
ma
rk
e
t
beha
v
iou
r
s
.
Rega
r
ding
t
he
la
tt
e
r,
and
despi
t
e
t
he
success
of
online
ensembles
dealing
w
i
t
h
complex
s
y
s
t
ems
and
t
r
aining
base
lea
r
ne
r
s
t
o
deal
w
i
t
h
diffe
r
en
t
r
egimes
,
t
he
use
of
t
hese
app
r
oaches
f
r
om
t
he
da
t
a
s
t
r
eam
lea
r
ning
li
t
e
r
a
t
u
r
e
is
no
t
as
popula
r
in
financial
fo
r
ecas
t
ing
y
e
t
.
E
v
en
t
hough
t
he
r
e
doesn’
t
seem
t
o
be
a
clea
r
l
y
dominan
t
ML
t
echnique
in
t
his
space
,
i
t
is
w
o
r
t
h
men
t
ioning
t
he
popula
r
i
t
y
of
solu
t
ions
based
on
ensembles
and
e
v
ol
v
ing
fuzz
y
s
y
s
t
ems
.
It
is
also
impo
r
t
an
t
t
o
no
t
e
ho
w
t
he
r
ela
t
i
v
el
y r
ecen
t
de
v
elopmen
t
s
in
deep
lea
r
ning
ha
v
e
fos
t
e
r
ed
t
he
popula
r
i
t
y
of
app
r
oaches
w
he
r
e
a
r
t
ificial
neu
r
al
ne
t
w
o
rk
s
pla
y
a
k
e
y r
ole
.
Fu
t
u
r
e
r
esea
r
ch
is
li
k
el
y
t
o
emphasise
t
he
applica
t
ion
of
da
t
a
s
t
r
eam
classifica
t
ion
algo
r
i
t
hms
t
o
financial
s
t
r
eams
.
Online
machine
lea
r
ning
has
no
t
been
w
idel
y
applied
t
o
t
he
financial
domain
. H
o
w
e
v
e
r,
as
sho
w
n
in
t
his
s
t
ud
y,
simila
r
t
echniques
li
k
e
sequen
t
ial
and
r
ecu
rr
ing
deep
lea
r
ning
models
a
r
e
on
t
he
r
ise
in
finance
. A
ppl
y
ing
t
he
p
r
oblem
of
concep
t
d
r
if
t
t
o
handling
p
r
ice
change
d
y
namics
seems
a
na
t
u
r
al
s
t
ep
fo
rw
a
r
d
on
t
he
r
esea
r
ch
line
of
financial
r
egime
changes
.
H
a
v
ing
said
t
ha
t
,
be
tt
e
r
access
t
o
high
-
f
r
equenc
y
da
t
a
and
compu
t
a
t
ional
r
esou
r
ces
w
ill
also
li
k
el
y r
esul
t
in
majo
r
p
r
og
r
ess
in
t
he
nea
r
fu
t
u
r
e
.
Ac
k
no
w
l
e
dgm
e
nt
W
e
w
ould
li
k
e
t
o
t
han
k
t
he
edi
t
o
r
and
ex
t
e
r
nal
r
e
v
ie
w
e
r
s
fo
r
t
hei
r
t
hough
t
ful
and
de
t
ailed
commen
t
s
on
ou
r
pape
r. W
e
w
ould
also
li
k
e
t
o
ac
k
no
w
ledge
t
he
financial
suppo
r
t
of
t
he
Spanish
M
inis
t
ry
of
Science
,
I
nno
v
a
t
ion
and
U
ni
v
e
r
si
t
ies
unde
r
g
r
an
t
P
GC
2018
-
096849
-
B
-
I
00
(
MC
Fin)
.
Re
f
e
r
e
nc
e
s
.
Y
r
[
1
]
Y
.
S
. A
bu
-M
os
t
afa
, A.
F
. A
t
i
y
a
,
I
n
t
r
oduc
t
ion
t
o
f
inancial
fo
r
ecas
t
in
g,
A
pp
li
ed
I
nte
llig
ence
, v
ol
.
6
,
pp
.
205–213
,
7
1996
.
[
2
]
W. H
uan
g,
Y
. N
a
k
amo
r
i
,
S
.-
Y
. W
an
g,
Fo
r
ecas
t
in
g
s
t
oc
k
ma
rk
e
t
mo
v
emen
t
di
r
ec
t
ion
w
i
t
h
suppo
r
t
v
ec
t
o
r
machine
, C
o
m
p
u
ter
s
&
Operat
i
on
s
Re
s
earc
h,
v
ol
.
32
,
pp
.
2513–2522
,
2005
.
[
3
]
R
. C. C
a
v
alcan
t
e
,
R
. C. Br
asilei
r
o
, V. L.
F
.
Souza
,
J
.
P
. N
ob
r
e
g
a
, A. L.
I
.
Oli
v
ei
r
a
, “C
ompu
t
a
t
ional
I
n
t
elli
g
ence
and
Financial
M
a
rk
e
t
s
: A
Su
rv
e
y
and
Fu
t
u
r
e
D
i
r
ec
t
ions
,
E
x
pert
S
y
s
te
m
s
wi
t
h A
pp
li
cat
i
on
s
, v
ol
.
55
,
pp
.
194–211
,
8
2016
.
[
4
]
V. D
o
gr
a
,
S
. V
e
r
ma
, K
a
v
i
t
a
, N. Z.
Jhanjhi
, U. G
hosh
, D. N. L
e
, A
C
ompa
r
a
t
i
v
e
A
nal
y
sis
of
M
achine
L
ea
r
nin
g M
odels
fo
r B
an
k
in
g N
e
w
s
E
x
t
r
ac
t
ion
b
y M
ul
t
iclass
C
lassi
f
ica
t
ion
W
i
t
h
I
mbalanced
D
a
t
ase
t
s
of
Financial
N
e
w
s
: C
hallen
g
es
and
Solu
t
ions
,
I
nternat
i
ona
l J
o
u
rna
l
o
f
I
nteract
iv
e
M
ul
t
im
ed
i
a
and
A
rt
i
f
i
c
i
a
l
I
nte
llig
ence
, v
ol
.
7
,
no
.
3
,
pp
.
35–52
,
2022
,
doi
:
10
.
9781/ijimai
.
2022
.
02
.
002
.
[
5
]
M. W. H
su
,
S
. L
essmann
, M. C.
Sun
g, T M
a
,
J
. E.
Johnson
, “Br
id
g
in
g
t
he
di
v
ide
in
f
inancial
ma
rk
e
t
fo
r
ecas
t
in
g:
machine
lea
r
ne
r
s
v
s
. f
inancial
economis
t
s
,
E
x
pert
S
y
s
te
m
s
wi
t
h A
pp
li
cat
i
on
s
, v
ol
.
61
,
pp
.
215–234
,
2016
.
[
6
]
B. M. H
en
r
ique
, V. A.
Sob
r
ei
r
o
, H. K
imu
r
a
, “L
i
t
e
r
a
t
u
r
e
r
e
v
ie
w: M
achine
lea
r
nin
g
t
echniques
applied
t
o
f
inancial
ma
rk
e
t
p
r
edic
t
ion
,
E
x
pert
S
y
s
te
m
s
wi
t
h A
pp
li
cat
i
on
s
, v
ol
.
124
,
pp
.
226–251
,
2019
.
[
7
]
G.
S
. A
t
sala
k
is
, K.
P
. V
ala
v
anis
,
Su
rv
e
y
in
g
s
t
oc
k
ma
rk
e
t
fo
r
ecas
t
in
g
t
echniques
-
Pa
r
t
II
:
Sof
t
compu
t
in
g
me
t
hods
,
E
x
pert
S
y
s
te
m
s
wi
t
h
A
pp
li
cat
i
on
s
, v
ol
.
36
,
pp
.
5932
-
59
-
41
,
4
2009
.
[
8
]
M.- . C
hen
, A. K.
San
g
aiah
, T.-H. C
hen
, E. D. L
u
g
hofe
r, E. Egr
io
g
lu
,
“D
eep
lea
r
nin
g
fo
r f
inancial
en
g
inee
r
in
g,C
o
m
p
u
tat
i
ona
l
Econo
mi
c
s
,
pp
.
1–5
,
2022
.
[
9
]
A. M.
Ozba
y
o
g
lu
, M. U. G
udele
k,
O
. B.
Seze
r, “D
eep
lea
r
nin
g
fo
r f
inancial
applica
t
ions
: A
su
rv
e
y,A
pp
li
ed
So
f
t
C
o
m
p
u
t
i
n
g, v
ol
.
93
,
p
.
106384
,
2020
,
doi
:
h
tt
ps
:
//doi
.
o
rg
/10
.
1016/j
.
asoc
.
2020
.
106384
.
[
10
]
M. B
allin
g
s
, D. V
an
D
en
Poel
, N. H
espeels
,
R
. Gry
p
, “Ev
alua
t
in
g
mul
t
iple
classi
f
ie
r
s
fo
r
s
t
oc
k
p
r
ice
di
r
ec
t
ion
p
r
edic
t
ion
,
E
x
pert
S
y
s
te
m
s
wi
t
h
A
pp
li
cat
i
on
s
, v
ol
.
42
,
no
.
20
,
pp
.
7046–7056
,
2015
.
[
11
]
A. B
oo
t
h
, E. G
e
r
din
g,
F
. M
c
Gr
oa
r
t
y, “A
u
t
oma
t
ed
t
r
adin
g w
i
t
h
pe
r
fo
r
mance
w
ei
g
h
t
ed
r
andom
fo
r
es
t
s
and
seasonali
t
y,
E
x
pert
S
y
s
te
m
s
wi
t
h A
pp
li
cat
i
on
s
, v
ol
.
41
,
pp
.
3651–3661
,
6
2014
.
[
12
]
P
. Ł
ad
y
ż
y
ńs
k
i
, K. Ż
bi
k
o
w
s
k
i
,
P
. Gr
ze
g
o
r
ze
w
s
k
i
,
S
t
oc
k
t
r
adin
g w
i
t
h
r
andom
fo
r
es
t
s
,
t
r
end
de
t
ec
t
ion
t
es
t
s
and
fo
r
ce
index
v
olume
indica
t
o
r
s
,
in
A
rt
i
f
i
c
i
a
l
I
nte
llig
ence
and
So
f
t
C
o
m
p
u
t
i
n
g:
12
t
h
I
nternat
i
ona
l C
on
f
erence
,
I
CA
I
S
C
2013
,
P
roceed
i
n
g
s
,
P
art
II
, B
e
r
lin
, H
eidelbe
rg,
Jun
.
2013
,
pp
.
441–
452
,
Sp
r
in
g
e
r B
e
r
lin
H
eidelbe
rg.
[
13
]
J
.
Pa
t
el
,
S
.
Shah
,
P
. T
ha
kk
a
, K. K
o
t
echa
, “
P
r
edic
t
in
g
s
t
oc
k
and
s
t
oc
k
p
r
ice
R
e
g
u
l
a
r
Iss
ue
-
1
4
5 -
r
r
index
mo
v
emen
t
usin
g
t
r
end
de
t
e
r
minis
t
ic
da
t
a
p
r
epa
r
a
t
ion
and
machine
lea
r
nin
g
t
echniques
,
E
x
pert
S
y
s
te
m
s
wi
t
h A
pp
li
cat
i
on
s
, v
ol
.
42
,
no
.
1
,
pp
.
259–268
,
2015
,
doi
:
h
tt
ps
:
//doi
.
o
rg
/10
.
1016/j
.
es
w
a
.
2014
.
07
.
040
.
[
14
]
D. Ar
dia
, K. B
lu
t
eau
, M.
Rüede
,
Re
g
ime
chan
g
es
in
B
i
t
coin
GA
R
CH
v
ola
t
ili
t
y
d
y
namics
,
F
i
nance
Re
s
earc
h L
etter
s
, v
ol
.
29
,
pp
.
266–271
,
Jun
.
2019
,
doi
:
10
.
1016/J
.
FR
L.
2018
.
08
.
009
.
[
15
]
A. A
n
g, A. T
imme
r
mann
,
Re
g
ime
chan
g
es
and
f
inancial
ma
rk
e
t
s
,
A
nn
u
a
l
Re
vi
e
w
o
f
F
i
nanc
i
a
l
Econo
mi
c
s
, v
ol
.
4
,
no
.
1
,
pp
.
313–337
,
2012
.
[
16
]
A. T
s
y
mbal
, “T
he
P
r
oblem
of
C
oncep
t
Dr
if
t
: D
e
f
ini
t
ions
and
Rela
t
ed
W
o
rk,
Tec
h
n
i
ca
l
Report
:
T
CD-C
S
-
2004
-
15
, D
epart
m
ent
o
f C
o
m
p
u
ter
Sc
i
ence
Tr
i
n
i
t
y
C
o
ll
e
g
e
, Du
b
li
n
,
2004
.
[
17
]
A. L.
Suá
r
ez
-C
e
t
r
ulo
, A. C
e
rv
an
t
es
, D.
Quin
t
ana
,
I
nc
r
emen
t
al
M
a
rk
e
t
B
eha
v
io
r C
lassi
f
ica
t
ion
in
P
r
esence
of
Recu
rr
in
g C
oncep
t
s
,
Entrop
y
, v
ol
.
21
,
p
.
25
,
Jan
.
2019
,
doi
:
10
.
3390/e21010025
.
[
18
]
M. C. M
ünnix
, T.
Shimada
,
R
.
Schäfe
r,
F
. L
e
yvr
az
, T. H.
Seli
g
man
, T. G
uh
r, H.
E.
S
t
anle
y, “
I
den
t
if
y
in
g
S
t
a
t
es
of
a
Financial
M
a
rk
e
t
,
Sc
i
ent
i
f
i
c
Report
s
, v
ol
.
2
,
p
.
644
,
12
2012
.
[
19
]
E. T
san
g,
J
. C
hen
, D
etect
i
n
g
re
gim
e
c
h
an
g
e
i
n
co
m
p
u
tat
i
ona
l
f
i
nance
:
data
s
c
i
ence
, m
ac
hi
ne
l
earn
i
n
g
and
a
lg
or
i
t
hmi
c
trad
i
n
g. C
R
C
P
r
ess
,
2020
.
[
20
]
R
. T. D
as
, K. K. A
n
g, C.
Que
k,
I
eRSPOP
: A
no
v
el
inc
r
emen
t
al
r
ou
g
h
se
t
-
based
pseudo
ou
t
e
r-
p
r
oduc
t
w
i
t
h
ensemble
lea
r
nin
g, A
pp
li
ed
So
f
t
C
o
m
p
u
t
i
n
g J
o
u
rna
l, v
ol
.
46
,
pp
.
170–186
,
9
2016
.
[
21
]
V. V
ella
, W. L. Ng, “E
nhancin
g r
is
k-
adjus
t
ed
pe
r
fo
r
mance
of
s
t
oc
k
ma
rk
e
t
in
t
r
ada
y
t
r
adin
g w
i
t
h
N
eu
r
o
-
Fuzz
y
s
y
s
t
ems
,
Ne
u
roco
m
p
u
t
i
n
g, v
ol
.
141
,
pp
.
170–187
,
2014
.
[
22
]
Y
. H
u
, K. L
iu
, X. Z
han
g, K. X
ie
, W. C
hen
,
Y
. Z
en
g, M. L
iu
, “C
oncep
t
d
r
if
t
minin
g
of
po
r
t
folio
selec
t
ion
fac
t
o
r
s
in
s
t
oc
k
ma
rk
e
t
,
E
l
ectron
i
c
C
o
mm
erce
Re
s
earc
h
and
A
pp
li
cat
i
on
s
, v
ol
.
14
,
no
.
6
,
pp
.
444–455
,
2015
.
[
23
]
B.
Sil
v
a
, N. M
a
r
ques
, G.
Panosso
, “A
ppl
y
in
g
neu
r
al
ne
t
w
o
rk
s
fo
r
concep
t
d
r
if
t
de
t
ec
t
ion
in
f
inancial
ma
rk
e
t
s
,
in
C
E
U
R
W
or
k
s
h
op
P
roceed
i
n
g
s
, v
ol
.
960
,
2012
,
pp
.
43–47
.
[
24
]
X. G
u
,
P
.
P
. A
n
g
elo
v, A. M. A
li
, W. A. Gr
u
v
e
r, G. G
a
y
dadjie
v,
Online
e
v
ol
v
in
g
fuzz
y r
ule
-
based
p
r
edic
t
ion
model
fo
r
hi
g
h
f
r
equenc
y
t
r
adin
g
f
inancial
da
t
a
s
t
r
eam
,
in
2016
I
EEE
C
on
f
erence
on
E
v
o
lvi
n
g
and
A
dapt
iv
e
I
nte
llig
ent
S
y
s
te
m
s
(E
A
I
S)
,
5
2016
,
pp
.
169–175
,
I
EEE.
[
25
]
E.
F
.
Fama
, “E
f
f
icien
t
C
api
t
al
M
a
rk
e
t
s
: A
Re
v
ie
w
of
T
heo
ry
and
E
mpi
r
ical
W
o
rk,
T
h
e
J
o
u
rna
l
o
f
F
i
nance
, v
ol
.
25
,
no
.
2
,
p
.
383
,
1970
,
doi
:
10
.
2307/2325486
.
[
26
]
J
.
Pi
g
e
r,
Econo
m
etr
i
c
s
:
Mode
l
s
o
f
Re
gim
e
Ch
an
g
e
s
,
pp
.
2744–2757
. N
e
w
Yo
rk, N
Y
:
Sp
r
in
g
e
r N
e
w
Yo
rk,
2009
.
[
27
]
A. G. H
oepne
, D. M
c
M
illan
, A. V
i
v
ian
, W
ese
Simen
,
Si
g
ni
f
icance
,
r
ele
v
ance
and
explainabili
t
y
in
t
he
machine
lea
r
nin
g
a
g
e
:
an
econome
t
r
ics
and
f
inancial
da
t
a
science
pe
r
spec
t
i
v
e
,
T
h
e
E
u
ropean
J
o
u
rna
l
o
f
F
i
nance
,
v
ol
.
27
,
no
.
1
-
2
,
pp
.
1–7
,
2021
.
[
28
]
P
. Br
ac
k
e
, A. D
a
tt
a
, C.
Jun
g,
S
.
Sen
, “M
achine
lea
r
nin
g
explainabili
t
y
in
f
inance
:
an
applica
t
ion
t
o
defaul
t
r
is
k
anal
y
sis
,” B
an
k
of
E
n
g
land
,
2019
.
[
29
]
M. Kr
i
t
zman
,
S
.
Pa
g
e
, D. T
u
rk
in
g
t
on
,
Re
g
ime
shif
t
s
:
I
mplica
t
ions
fo
r
d
y
namic
s
t
r
a
t
e
g
ies
(co
rr
ec
t
ed)
,
F
i
nanc
i
a
l A
na
l
y
s
t
s
J
o
u
rna
l, v
ol
.
68
,
no
.
3
,
pp
.
22–39
,
2012
.
[
30
]
R
. E
l
w
ell
,
R
.
Poli
k
a
,
I
nc
r
emen
t
al
lea
r
nin
g
of
concep
t
d
r
if
t
in
nons
t
a
t
iona
ry
en
v
i
r
onmen
t
s
.,
I
EEE
tran
s
act
i
on
s
on
ne
u
ra
l
net
w
or
k
s
, v
ol
.
22
,
pp
.
1517–31
,
10
2011
.
[
31
]
J
. A. G
ama
,
I
. Ž
liobai
t
unde
f
ined
, A. B
ife
t
, M.
Pecheniz
k
i
y, A. B
ouchachia
, A
su
rv
e
y
on
concep
t
d
r
if
t
adap
t
a
t
ion
,AC
M
C
o
m
p
u
t
i
n
g
S
u
r
v
e
y
s
, v
ol
.
46
, M
a
r.
2014
,
doi
:
10
.
1145/2523813
.
[
32
]
G. D
i
t
zle
r, M.
Ro
v
e
r
i
, C. A
lippi
,
R
.
Poli
k
a
r, “L
ea
r
nin
g
in
nons
t
a
t
iona
ry
en
v
i
r
onmen
t
s
: A
su
rv
e
y,
I
EEE
C
o
m
p
u
tat
i
ona
l
I
nte
llig
ence
Ma
g
a
zi
ne
, v
ol
.
10
,
pp
.
12–25
, N
o
v.
2015
.
[
33
]
G.
I
. W
ebb
,
R
. Hy
de
, H. C
ao
, H. L. Ng
u
y
en
,
F
.
Pe
t
i
t
jean
, “C
ha
r
ac
t
e
r
izin
g
concep
t
d
r
if
t
, D
ata
M
i
n
i
n
g
and
K
no
wl
ed
g
e
Di
s
co
v
er
y
, v
ol
.
30
,
pp
.
964–
994
,
7
2016
.
[
34
]
H. M. G
omes
,
J
.
P
. B
a
r
ddal
,
F
. E
nemb
r
ec
k, A. B
ife
t
, A
Su
rv
e
y
on
E
nsemble
L
ea
r
nin
g
fo
r D
a
t
a
S
t
r
eam
C
lassi
f
ica
t
ion
, AC
M
C
o
m
p
u
t
i
n
g
S
u
r
v
e
y
s
, v
ol
.
50
,
no
.
2
,
pp
.
1–36
,
2017
,
doi
:
10
.
1145/3054925
.
[
35
]
S
.
Ramí
r
ez
-G
alle
g
o
, B. Kr
a
w
cz
yk,
S
. G
a
r
cía
, M. W
oźnia
k,
F
. H
e
rr
e
r
a
, A
su
rv
e
y
on
da
t
a
p
r
ep
r
ocessin
g
fo
r
da
t
a
s
t
r
eam
minin
g: C
u
rr
en
t
s
t
a
t
us
and
fu
t
u
r
e
di
r
ec
t
ions
,
Ne
u
roco
m
p
u
t
i
n
g, v
ol
.
239
,
pp
.
39–57
,
2017
.
[
36
]
A.
R
. M
ase
g
osa
, A. M. M
a
r
t
ínez
, D.
Ramos
-L
ópez
, H. L
an
g
se
t
h
, T. D.
N
ielsen
, A.
Salme
r
ón
, “A
nal
y
zin
g
concep
t
d
r
if
t
: A
case
s
t
ud
y
in
t
he
f
inancial
sec
t
o
r,
I
nte
llig
ent
D
ata
A
na
l
y
s
i
s
, v
ol
.
24
,
no
.
3
,
pp
.
665–688
,
2020
.
,
Y
[
37
]
M.
P
r
a
t
ama
, E. L
u
g
hofe
r,
J
. Er,
S
. A
na
v
a
tt
i
, C.-
P
. L
im
, “D
a
t
a
d
r
i
v
en
modellin
g
based
on
Recu
rr
en
t
I
n
t
e
rv
al
-V
alued
M
e
t
aco
g
ni
t
i
v
e
Scaffoldin
g
Fuzz
y N
eu
r
al
N
e
t
w
o
rk,
Ne
u
roco
m
p
u
t
i
n
g, v
ol
.
262
,
pp
.
4–27
,
2017
.
[
38
]
M.
P
r
a
t
ama
,
J
. L
u
, E. L
u
g
hofe
r, G. Z
han
g, M.
J
. Er
I
nc
r
emen
t
al
L
ea
r
nin
g
of
C
oncep
t
Dr
if
t
U
sin
g Ev
ol
v
in
g Ty
pe
-
2
Recu
rr
en
t
Fuzz
y N
eu
r
al
N
e
t
w
o
rk,
I
EEE
Tran
s
act
i
on
s
on
F
uzz
y
S
y
s
te
m
s
,
pp
.
1–1
,
2016
,
doi
:
10
.
1109/
T
F
UZZ.
2016
.
2599855
.
[
39
]
C. A
lippi
, G. B
o
r
acchi
, M.
Ro
v
e
r
i
,
Jus
t
-
in
-
t
ime
classi
f
ie
r
s
fo
r r
ecu
rr
en
t
concep
t
s
,
I
EEE
Tran
s
act
i
on
s
on
Ne
u
ra
l
Net
w
or
k
s
and
L
earn
i
n
g
S
y
s
te
m
s
,
v
ol
.
24
,
pp
.
620–634
,
4
2013
.
[
40
]
J
. B. G
omes
, M. M. G
abe
r,
P
. A. C.
Sousa
, E. M
enasal
v
as
, “M
inin
g
r
ecu
rr
in
g
concep
t
s
in
a
d
y
namic
fea
t
u
r
e
space
,
I
EEE
Tran
s
act
i
on
s
on
Ne
u
ra
l
Net
w
or
k
s
and
L
earn
i
n
g
S
y
s
te
m
s
, v
ol
.
25
,
pp
.
95–110
,
1
2014
,
doi
:
10
.
1109/
TNNL
S
.
2013
.
2271915
.
[
41
]
P
. M. G
onçal
v
es
J
r,
R
.
Sou
t
o
, M. D
e
B
a
rr
os
,
R
CD: A r
ecu
rr
in
g
concep
t
d
r
if
t
f
r
ame
w
o
rk,
P
attern
Reco
g
n
i
t
i
on
L
etter
s
, v
ol
.
34
,
pp
.
1018–1025
,
2013
.
[
42
]
A. W. L
o
, H. M
ama
y
s
ky,
J
. W
an
g,
Founda
t
ions
of
T
echnical
A
nal
y
sis
:
C
ompu
t
a
t
ional
A
l
g
o
r
i
t
hms
,
S
t
a
t
is
t
ical
I
nfe
r
ence
,
and
E
mpi
r
ical
I
mplemen
t
a
t
ion
,
T
h
e
J
o
u
rna
l
o
f
F
i
nance
, v
ol
.
55
,
pp
.
1705–1765
,
8
2000
.
[
43
]
F
. E. T
a
y, L. C
ao
, “A
pplica
t
ion
of
suppo
r
t
v
ec
t
o
r
machines
in
f
inancial
t
ime
se
r
ies
fo
r
ecas
t
in
g,
O
m
e
g
a
, v
ol
.
29
,
pp
.
309–317
,
8
2001
.
[
44
]
T. G
e
v
a
,
J
. Z
aha
v
i
, “E
mpi
r
ical
e
v
alua
t
ion
of
an
au
t
oma
t
ed
in
t
r
ada
y
s
t
oc
k
r
ecommenda
t
ion
s
y
s
t
em
inco
r
po
r
a
t
in
g
bo
t
h
ma
rk
e
t
da
t
a
and
t
ex
t
ual
ne
w
s
, D
ec
i
s
i
on
S
u
pport
S
y
s
te
m
s
, v
ol
.
57
,
pp
.
212–223
,
Jan
.
2014
,
doi
:
10
.
1016/J
.D
SS
.
2013
.
09
.
013
.
[
45
]
C. H. C
hen
,
P
. . C
hen
,
J
. C. W. L
in
, A
n
E
nsemble
C
lassi
f
ie
r
fo
r
S
t
oc
k
Tr
end
P
r
edic
t
ion
U
sin
g
Sen
t
ence
- L
e
v
el
C
hinese
N
e
w
s
Sen
t
imen
t
and
T
echnical
I
ndica
t
o
r
s
,
I
nternat
i
ona
l J
o
u
rna
l
o
f
I
nteract
iv
e
M
ul
t
im
ed
i
a
and
A
rt
i
f
i
c
i
a
l
I
nte
llig
ence
, v
ol
.
7
,
no
.
3
,
pp
.
53–64
,
2022
,
doi
:
10
.
9781/
ijimai
.
2022
.
02
.
004
.
[
46
]
B. K
i
t
chenham
,
O
.
P
. Br
e
r
e
t
on
, D. B
ud
g
en
, M. T
u
r
ne
r,
J
. B
aile
y,
S
.
L
in
k
man
,
S
y
s
t
ema
t
ic
li
t
e
r
a
t
u
r
e
r
e
v
ie
w
s
in
sof
t
w
a
r
e
en
g
inee
r
in
g
–a
s
y
s
t
ema
t
ic
li
t
e
r
a
t
u
r
e
r
e
v
ie
w,
I
n
f
or
m
at
i
on
and
s
o
f
t
w
are
tec
h
no
l
o
g
y
, v
ol
.
51
,
no
.
1
,
pp
.
7–
15
,
2009
.
[
47
]
B. K
i
t
chenham
,
P
r
ocedu
r
es
fo
r
pe
r
fo
r
min
g
s
y
s
t
ema
t
ic
r
e
v
ie
w
s
, K
ee
l
e
,
UK, K
ee
l
e
U
n
iv
er
s
i
t
y
, v
ol
.
33
,
no
.
2004
,
pp
.
1–26
,
2004
.
[
48
]
M.
J
.
Pa
g
e
,
J
. E. M
c
K
enzie
,
P
. M. B
ossu
y
t
,
I
. B
ou
t
r
on
, T. C. H
offmann
, C. D.
M
ul
r
o
w, L.
Shamsee
r,
J
. M. T
e
t
zlaff
, E. A. Ak
l
,
S
. E. Br
ennan
,
et
a
l
.
, “T
he
p
r
isma
2020
s
t
a
t
emen
t
:
an
upda
t
ed
g
uideline
fo
r r
epo
r
t
in
g
s
y
s
t
ema
t
ic
r
e
v
ie
w
s
,
S
y
s
te
m
at
i
c
re
vi
e
w
s
, v
ol
.
10
,
no
.
1
,
pp
.
1–11
,
2021
.
[
49
]
A. W. L
o
,
Reconcilin
g
ef
f
icien
t
ma
rk
e
t
s
w
i
t
h
beha
v
io
r
al
f
inance
:
t
he
adap
t
i
v
e
ma
rk
e
t
s
h
y
po
t
hesis
,J
o
u
rna
l
o
f i
n
v
e
s
t
m
ent
con
s
ul
t
i
n
g, v
ol
.
7
,
no
.
2
,
pp
.
21–44
,
2005
.
[
50
]
S
. B
ae
k,
S
. K. M
ohan
t
y, M. G
lambos
ky, “C
o
v
id
-
19
and
s
t
oc
k
ma
rk
e
t
v
ola
t
ili
t
y: A
n
indus
t
ry
le
v
el
anal
y
sis
,
F
i
nance
Re
s
earc
h L
etter
s
, v
ol
.
37
,
p
.
101748
,
2020
.
[
51
]
M. L. D
e
P
r
ado
, A
d
v
ance
s
i
n
f
i
nanc
i
a
l m
ac
hi
ne
l
earn
i
n
g.
John
W
ile
y
&
Sons
,
2018
.
[
52
]
A. A
n
g, G. B
e
k
ae
r
t
, “H
o
w r
e
g
imes
affec
t
asse
t
alloca
t
ion
,
F
i
nanc
i
a
l
A
na
l
y
s
t
s
J
o
u
rna
l, v
ol
.
60
,
no
.
2
,
pp
.
86–99
,
2004
.
[
53
]
G. D
a
v
ies
,
Re
g
ime
chan
g
es
in
t
he
f
inancial
ma
rk
e
t
s
,
2016
.
[
Online
]
. Av
ailable
:
h
tt
ps
:
//
www.
f
t
.
com/con
t
en
t
/6556ec60
-
6aa3
-
3dfe
-
8953
-
b94d6080c360
.
[
54
]
E. A
nd
r
eou
, E. G
h
y
sels
,
Str
u
ct
u
ra
l
Brea
k
s
i
n
F
i
nanc
i
a
l
T
im
e
Ser
i
e
s
,
pp
.
839–870
. B
e
r
lin
, H
eidelbe
rg:
Sp
r
in
g
e
r B
e
r
lin
H
eidelbe
rg,
2009
.
[
55
]
A.
Sami
t
as
, E. K
ampou
r
is
, D. K
enou
rg
ios
, “M
achine
lea
r
nin
g
as
an
ea
r
l
y w
a
r
nin
g
s
y
s
t
em
t
o
p
r
edic
t
f
inancial
c
r
isis
,
I
nternat
i
ona
l
Re
vi
e
w
o
f
F
i
nanc
i
a
l A
na
l
y
s
i
s
, v
ol
.
71
,
p
.
101507
,
2020
.
[
56
]
D.
Pe
tt
enuzzo
, A. T
imme
r
mann
, “
P
r
edic
t
abili
t
y
of
s
t
oc
k r
e
t
u
r
ns
and
asse
t
alloca
t
ion
unde
r
s
t
r
uc
t
u
r
al
b
r
ea
k
s
,J
o
u
rna
l
o
f
Econo
m
etr
i
c
s
, v
ol
.
164
,
pp
.
60–78
,
Sep
.
2011
,
doi
:
10
.
1016/j
.
jeconom
.
2011
.
02
.
019
.
[
57
]
A. W. L
o
, “T
he
A
dap
t
i
v
e
M
a
rk
e
t
s
Hy
po
t
hesis
: M
a
rk
e
t
E
f
f
icienc
y
f
r
om
an
Ev
olu
t
iona
ry
Pe
r
spec
t
i
v
e
,” M
I
T,
2004
.
[
58
]
J
. D. H
amil
t
on
, A
ne
w
app
r
oach
t
o
t
he
economic
anal
y
sis
of
nons
t
a
t
iona
ry
t
ime
se
r
ies
and
t
he
business
c
y
cle
,
Econo
m
etr
i
ca
: J
o
u
rna
l
o
f
t
h
e
econo
m
etr
i
c
s
oc
i
et
y
,
pp
.
357–384
,
1989
.
[
59
]
J
. D. H
amil
t
on
,
Re
g
ime
s
w
i
t
chin
g
models
,
in
Macroecono
m
etr
i
c
s
and
t
im
e
s
er
i
e
s
ana
l
y
s
i
s
,
Sp
r
in
g
e
r,
2010
,
pp
.
202–209
.
[
60
]
T.
P
r
eis
,
J
.
J
.
Schneide
r, H. E.
S
t
anle
y,
S
w
i
t
chin
g
p
r
ocesses
in
f
inancial
ma
rk
e
t
s
,
P
roceed
i
n
g
s
o
f
t
h
e
Nat
i
ona
l A
cade
m
y
o
f
Sc
i
ence
s
, v
ol
.
108
,
no
.
19
,
pp
.
7674–7678
,
2011
,
doi
:
10
.
1073/pnas
.
1019484108/
-
/
DC
Supplemen
t
al
.
I
ntern
a
t
i
on
a
l
J
ourn
a
l
of
I
nter
ac
t
i
v
e
M
u
l
t
i
med
i
a a
nd
Art
i
f
i
c
i
a
l
I
nte
lli
g
en
c
e,
Vo
l
.
9
,
N
º
1
-
1
46
-
T
[
61
]
R
. H
amme
r
schmid
, H. L
oh
r
e
,
Re
g
ime
shif
t
s
and
s
t
oc
k r
e
t
u
r
n
p
r
edic
t
abili
t
y,
I
nternat
i
ona
l
Re
vi
e
w
o
f
Econo
mi
c
s
and
F
i
nance
, v
ol
.
56
,
pp
.
138–160
,
Jul
.
2018
,
doi
:
10
.
1016/j
.
i
r
ef
.
2017
.
10
.
021
.
[
62
]
Q
. D
ai
, K.
J
.
Sin
g
le
t
on
, W.
Yan
g,
I
s
r
e
g
ime
-
shif
t
r
is
k
p
r
iced
in
t
he
us
t
r
easu
ry
ma
rk
e
t
?
, W
o
rk
in
g
pape
r, N
e
w
Yo
rk U
ni
v
e
r
si
t
y
and
S
t
anfo
r
d
U
ni
v
e
r
si
t
y,
2003
.
[
63
]
J
. D. H
amil
t
on
, “M
ac
r
oeconomic
r
e
g
imes
and
r
e
g
ime
shif
t
s
,
Handboo
k
o
f
m
acroecono
mi
c
s
, v
ol
.
2
,
pp
.
163–
201
,
2016
.
[
64
]
J
. G. D
ias
,
J
. K. V
e
r
mun
t
,
S
.
Ramos
, “C
lus
t
e
r
in
g f
inancial
t
ime
se
r
ies
: N
e
w
insi
g
h
t
s
f
r
om
an
ex
t
ended
hidden
ma
rk
o
v
model
,
E
u
ropean
J
o
u
rna
l
o
f
Operat
i
ona
l
Re
s
earc
h, v
ol
.
243
,
no
.
3
,
pp
.
852–864
,
2015
.
[
65
]
K.-
S
. K
im
,
I
. H
an
, “T
he
clus
t
e
r-
indexin
g
me
t
hod
fo
r
case
-
based
r
easonin
g
usin
g
self
-
o
rg
anizin
g
maps
and
lea
r
nin
g v
ec
t
o
r
quan
t
iza
t
ion
fo
r
bond
r
a
t
in
g
cases
,
E
x
pert
s
y
s
te
m
s
wi
t
h
app
li
cat
i
on
s
, v
ol
.
21
,
no
.
3
,
pp
.
147–
156
,
2001
.
[
66
]
R
. B
isoi
,
P
. K. D
ash
, A
h
y
b
r
id
e
v
olu
t
iona
ry
d
y
namic
neu
r
al
ne
t
w
o
rk
fo
r
s
t
oc
k
ma
rk
e
t
t
r
end
anal
y
sis
and
p
r
edic
t
ion
usin
g
unscen
t
ed
k
alman
f
il
t
e
r,
A
pp
li
ed
So
f
t
C
o
m
p
u
t
i
n
g, v
ol
.
19
,
pp
.
41–56
,
2014
.
[
67
]
M. G
uidolin
, A. T
imme
r
mann
, A
n
econome
t
r
ic
model
of
nonlinea
r
d
y
namics
in
t
he
join
t
dis
t
r
ibu
t
ion
of
s
t
oc
k
and
bond
r
e
t
u
r
ns
, J
o
u
rna
l
o
f
A
pp
li
ed
Econo
m
etr
i
c
s
, v
ol
.
21
,
pp
.
1–22
,
Jan
.
2006
,
doi
:
10
.
1002/jae
.
824
.
[
68
]
A. Ur
quha
r
t
,
F
. M
c
Gr
oa
r
t
y, Ar
e
s
t
oc
k
ma
rk
e
t
s
r
eall
y
ef
f
icien
t
?
e
v
idence
of
t
he
adap
t
i
v
e
ma
rk
e
t
h
y
po
t
hesis
,
I
nternat
i
ona
l
Re
vi
e
w
o
f
F
i
nanc
i
a
l A
na
l
y
s
i
s
, v
ol
.
47
,
pp
.
39–49
,
2016
,
doi
:
h
tt
ps
:
//doi
.
o
rg
/10
.
1016/j
.
i
r
fa
.
2016
.
06
.
011
.
[
69
]
N. G.
Pa
v
lidis
, V.
P
.
Pla
g
iana
k
os
, D. K. T
asoulis
, M. N. Vr
aha
t
is
, “
Financial
fo
r
ecas
t
in
g
t
h
r
ou
g
h
unsupe
rv
ised
clus
t
e
r
in
g
and
neu
r
al
ne
t
w
o
rk
s
,
Operat
i
ona
l
Re
s
earc
h, v
ol
.
6
,
no
.
2
,
pp
.
103–127
,
2006
.
[
70
]
A. A. Ar
i
y
o
, A.
O
. A
de
w
umi
, C. K. Ay
o
, “
S
t
oc
k
p
r
ice
p
r
edic
t
ion
usin
g
t
he
A
R
I
MA
model
,
in
2014
UK
S
im-A
MSS
16
t
h
I
nternat
i
ona
l C
on
f
erence
on
C
o
m
p
u
ter
Mode
lli
n
g
and
S
imul
at
i
on
,
2014
,
pp
.
106–112
,
I
EEE.
[
71
]
S
. C
houdhu
ry,
S
. G
hosh
, A. B
ha
tt
acha
ry
a
, K.
J
.
Fe
r
nandes
, M. K. T
i
w
a
r
i
, A
r
eal
t
ime
clus
t
e
r
in
g
and
S
VM
based
p
r
ice
-v
ola
t
ili
t
y
p
r
edic
t
ion
fo
r
op
t
imal
t
r
adin
g
s
t
r
a
t
e
gy,
Ne
u
roco
m
p
u
t
i
n
g, v
ol
.
131
,
pp
.
419–
426
,
5
2014
.
[
72
]
S
.-H.
Pa
rk,
J
.-H. L
ee
,
J
.-W.
Son
g, T.-
S
.
Pa
rk, “
Fo
r
ecas
t
in
g
chan
g
e
di
r
ec
t
ions
fo
r f
inancial
t
ime
se
r
ies
usin
g
hidden
M
a
rk
o
v
model
,
in
I
nternat
i
ona
l
C
on
f
erence
on
Ro
ugh
Set
s
and
K
no
wl
ed
g
e
Tec
h
no
l
o
g
y
,
2009
,
pp
.
184–191
,
Sp
r
in
g
e
r.
[
73
]
J
.
Si
r
i
g
nano
,
R
. C
on
t
, “U
ni
v
e
r
sal
fea
t
u
r
es
of
p
r
ice
fo
r
ma
t
ion
in
f
inancial
ma
rk
e
t
s
:
pe
r
spec
t
i
v
es
f
r
om
deep
lea
r
nin
g,
Q
u
ant
i
tat
iv
e
F
i
nance
,
2019
,
doi
:
10
.
1080/14697688
.
2019
.
1622295
.
[
74
]
G. D
eboec
k,
Trad
i
n
g
on
t
h
e
ed
g
e
:
ne
u
ra
l, g
enet
i
c
,
and
fuzz
y
s
y
s
te
m
s
f
or
c
h
aot
i
c
f
i
nanc
i
a
l m
ar
k
et
s
. W
ile
y,
1994
.
[
75
]
S
.
Schulmeis
t
e
r,
P
r
o
f
i
t
abili
t
y
of
t
echnical
s
t
oc
k
t
r
adin
g: H
as
i
t
mo
v
ed
f
r
om
dail
y
t
o
in
t
r
ada
y
da
t
a?
,
Re
vi
e
w
o
f
F
i
nanc
i
a
l
Econo
mi
c
s
, v
ol
.
18
,
no
.
4
,
pp
.
190–
201
,
2009
,
doi
:
10
.
1016/j
.r
fe
.
2008
.
10
.
001
.
[
76
]
T.
Shin
t
a
t
e
, L.
Pichl
, “Tr
end
P
r
edic
t
ion
C
lassi
f
ica
t
ion
fo
r H
i
g
h
F
r
equenc
y
B
i
t
coin
T
ime
Se
r
ies
w
i
t
h
D
eep
L
ea
r
nin
g, J
o
u
rna
l
o
f
R
i
s
k
and
F
i
nanc
i
a
l
Mana
g
e
m
ent
, v
ol
.
12
,
p
.
17
,
Jan
.
2019
,
doi
:
10
.
3390/j
r
fm12010017
.
[
77
]
B. G. M
al
k
iel
, “T
he
ef
f
icien
t
ma
rk
e
t
h
y
po
t
hesis
and
i
t
s
c
r
i
t
ics
, J
o
u
rna
l
o
f
econo
mi
c
per
s
pect
iv
e
s
, v
ol
.
17
,
no
.
1
,
pp
.
59–82
,
2003
.
[
78
]
A. W. L
o
, A. C. M
ac
K
inla
y, A
Non
-
Rando
m W
a
lk D
o
w
n
W
a
ll
Street
.
P
r
ince
t
on
U
ni
v
e
r
si
t
y
P
r
ess
, D
ec
.
2011
.
[
79
]
X. D
in
g,
Y
. Z
han
g, . L
iu
,
J
. D
uan
, “U
sin
g
s
t
r
uc
t
u
r
ed
e
v
en
t
s
t
o
p
r
edic
t
s
t
oc
k
p
r
ice
mo
v
emen
t
: A
n
empi
r
ical
in
v
es
t
i
g
a
t
ion
,
in
P
roceed
i
n
g
s
o
f
t
h
e
2014
C
on
f
erence
on
E
m
p
i
r
i
ca
l
Met
h
od
s
i
n
Nat
u
ra
l L
an
gu
a
g
e
P
roce
ss
i
n
g
(EMN
L
P
)
,
2014
,
pp
.
1415–1425
.
[
80
]
F
. B
lac
k, “N
oise
,
T
h
e
J
o
u
rna
l
o
f
F
i
nance
, v
ol
.
41
,
529–543
,
Jul
.
1986
,
doi
:
10
.
1111/j
.
1540
-
6261
.
1986
.
t
b04513
.
x
.
[
81
]
B. M
al
k
iel
, A
Rando
m W
a
lk D
o
w
n
W
a
ll
Street
. WW N
o
r
t
on
&
C
ompan
y,
1973
.
[
82
]
B. V
ans
t
one
, G.
Finnie
, A
n
empi
r
ical
me
t
hodolo
gy
fo
r
de
v
elopin
g
s
t
oc
k
ma
rk
e
t
t
r
adin
g
s
y
s
t
ems
usin
g
a
r
t
i
f
icial
neu
r
al
ne
t
w
o
rk
s
,
E
x
pert
S
y
s
te
m
s
wi
t
h A
pp
li
cat
i
on
s
, v
ol
.
36
,
no
.
3
,
Pa
r
t
2
,
pp
.
6668–6680
,
2009
,
doi
:
h
tt
ps
:
//doi
.
o
rg
/10
.
1016/j
.
es
w
a
.
2008
.
08
.
019
.
[
83
]
S
. C
a
r
t
a
, A. C
o
rr
i
g
a
, A.
Fe
rr
ei
r
a
, A.
S
.
Podda
, D.
R
.
Recupe
r
o
, A
mul
t
i
-
la
y
e
r
and
mul
t
i
-
ensemble
s
t
oc
k
t
r
ade
r
usin
g
deep
lea
r
nin
g
and
deep
r
einfo
r
cemen
t
lea
r
nin
g,A
pp
li
ed
I
nte
llig
ence
, v
ol
.
51
,
no
.
2
,
pp
.
889–905
,
2021
.
[
84
]
C. M
a
r
t
ín
, D.
Quin
t
ana
,
P
.
I
sasi
, “Gr
amma
t
ical
e
v
olu
t
ion
-
based
ensembles
fo
r
al
g
o
r
i
t
hmic
t
r
adin
g,A
pp
li
ed
So
f
t
C
o
m
p
u
t
i
n
g, v
ol
.
84
,
p
.
105713
,
2019
,
doi
:
h
tt
ps
:
//doi
.
o
rg
/10
.
1016/j
.
asoc
.
2019
.
105713
.
[
85
]
C. C
hen
, W. D
on
g
xin
g, H. C
hun
y
an
,
Y
. X
iaojie
, “E
xploi
t
in
g
social
media
fo
r
s
t
oc
k
ma
rk
e
t
p
r
edic
t
ion
w
i
t
h
fac
t
o
r
iza
t
ion
machine
,
in
2014
I
EEE
/
W
I
C
/
AC
M
I
nternat
i
ona
l J
o
i
nt
C
on
f
erence
s
on
W
eb
I
nte
llig
ence
(
W
I
)
and
I
nte
llig
ent
Ag
ent
Tec
h
no
l
o
gi
e
s
(
I
A
T)
, v
ol
.
2
,
2014
,
pp
.
142–149
,
I
EEE.
[
86
]
M.
R
. H
assan
, K.
Ramamohana
r
ao
,
J
. K
am
r
uzzaman
, M.
Rahman
, M. M.
H
ossain
, A
hmm
-
based
adap
t
i
v
e
fuzz
y
infe
r
ence
s
y
s
t
em
fo
r
s
t
oc
k
ma
rk
e
t
fo
r
ecas
t
in
g,
Ne
u
roco
m
p
u
t
i
n
g, v
ol
.
104
,
pp
.
10–25
,
2013
.
[
87
]
X. Z
han
g,
Y
. L
i
,
S
. W
an
g, B.
Fan
g,
S
.
Y
.
Philip
, “E
nhancin
g
s
t
oc
k
ma
rk
e
t
p
r
edic
t
ion
w
i
t
h
ex
t
ended
coupled
hidden
M
a
rk
o
v
model
o
v
e
r
mul
t
i
-
sou
r
ced
da
t
a
, K
no
wl
ed
g
e
and
I
n
f
or
m
at
i
on
S
y
s
te
m
s
, v
ol
.
61
,
no
.
2
,
pp
.
1071–1090
,
2019
.
[
88
]
N. G
â
r
leanu
, L. H.
Pede
r
sen
, “E
f
f
icien
t
l
y
inef
f
icien
t
ma
rk
e
t
s
fo
r
asse
t
s
and
asse
t
mana
g
emen
t
,
T
h
e
J
o
u
rna
l
o
f
F
i
nance
, v
ol
.
73
,
no
.
4
,
pp
.
1663–
1712
,
2018
.
[
89
]
G. C.
F
r
iesen
,
P
. A. W
elle
r, L. M. D
unham
,
P
r
ice
t
r
ends
and
pa
tt
e
r
ns
in
t
echnical
anal
y
sis
: A
t
heo
r
e
t
ical
and
empi
r
ical
examina
t
ion
,J
o
u
rna
l
o
f
Ban
ki
n
g
and
F
i
nance
, v
ol
.
33
,
pp
.
1089–1100
,
Jun
.
2009
,
doi
:
10
.
1016/j
.
jban
kf
in
.
2008
.
12
.
010
.
[
90
]
C.
F
. L
iu
, C.
Y
.
Yeh
,
S
.
J
. L
ee
, “A
pplica
t
ion
of
t
y
pe
-
2
neu
r
o
-
fuzz
y
modelin
g
in
s
t
oc
k
p
r
ice
p
r
edic
t
ion
,A
pp
li
ed
So
f
t
C
o
m
p
u
t
i
n
g J
o
u
rna
l, v
ol
.
12
,
no
.
4
,
pp
.
1348–
1358
,
2012
,
doi
:
10
.
1016/j
.
asoc
.
2011
.
11
.
028
.
[
91
]
R
. C. C
a
v
alcan
t
e
, A. L.
Oli
v
ei
r
a
, A
n
au
t
onomous
t
r
ade
r
a
g
en
t
fo
r
t
he
s
t
oc
k
ma
rk
e
t
based
on
online
sequen
t
ial
ex
t
r
eme
lea
r
nin
g
machine
ensemble
,
in
2014
I
nternat
i
ona
l J
o
i
nt
C
on
f
erence
on
Ne
u
ra
l
Net
w
or
k
s
(
I
JC
NN)
,
2014
,
pp
.
1424–1431
,
I
EEE.
[
92
]
R
. G
a
r
cia
,
R
. L
u
g
e
r, E.
Renaul
t
, “E
mpi
r
ical
assessmen
t
of
an
in
t
e
r
t
empo
r
al
op
t
ion
p
r
icin
g
model
w
i
t
h
la
t
en
t
v
a
r
iables
,J
o
u
rna
l
o
f
Econo
m
etr
i
c
s
, v
ol
.
116
,
no
.
1
-
2
,
pp
.
49–83
,
2003
.
[
93
]
R
. D
ash
,
P
. K. D
ash
,
R
. B
isoi
, A
self
adap
t
i
v
e
diffe
r
en
t
ial
ha
r
mon
y
sea
r
ch
based
op
t
imized
ex
t
r
eme
lea
r
nin
g
machine
fo
r f
inancial
t
ime
se
r
ies
p
r
edic
t
ion
,
S
w
ar
m
and
E
v
o
lu
t
i
onar
y
C
o
m
p
u
tat
i
on
, v
ol
.
19
,
pp
.
25–
42
,
2014
.
[
94
]
A. A. A
debi
y
i
, C. K. Ay
o
, M.
O
. A
debi
y
i
,
S
.
O
.
O
t
o
k
i
t
i
,
S
t
oc
k
p
r
ice
p
r
edic
t
ion
usin
g
neu
r
al
ne
t
w
o
rk w
i
t
h
h
y
b
r
idized
ma
rk
e
t
indica
t
o
r
s
,
J
o
u
rna
l
o
f
E
m
er
gi
n
g
Trend
s
i
n
C
o
m
p
u
t
i
n
g
and
I
n
f
or
m
at
i
on
Sc
i
ence
s
, v
ol
.
3
,
no
.
1
,
pp
.
1–9
,
2012
.
[
95
]
R
. T
saih
,
Y
. H
su
, C. C. L
ai
, “
Fo
r
ecas
t
in
g
S
&
P
500
s
t
oc
k
index
fu
t
u
r
es
w
i
t
h
a
h
y
b
r
id
A
I
s
y
s
t
em
,D
ec
i
s
i
on
S
u
pport
S
y
s
te
m
s
, v
ol
.
23
,
pp
.
161–174
,
6
1998
.
[
96
]
J
. C. N
e
v
es
, A. V
iei
r
a
,
I
mp
r
o
v
in
g
ban
kr
up
t
c
y
p
r
edic
t
ion
w
i
t
h
hidden
la
y
e
r
lea
r
nin
g v
ec
t
o
r
quan
t
iza
t
ion
,
E
u
ropean
A
cco
u
nt
i
n
g
Re
vi
e
w, v
ol
.
15
,
no
.
2
,
pp
.
253–271
,
2006
.
[
97
]
Y
. L.
Yon
g,
Y
. L
ee
, X. G
u
,
P
.
P
. A
n
g
elo
v, D. C. L. Ng
o
, E.
Sha
f
ipou
r, “
Fo
r
ei
g
n
cu
rr
enc
y
exchan
g
e
r
a
t
e
p
r
edic
t
ion
usin
g
neu
r
o
-
fuzz
y
s
y
s
t
ems
,
P
roced
i
a
co
m
p
u
ter
s
c
i
ence
, v
ol
.
144
,
pp
.
232–238
,
2018
.
[
98
]
H. G
hoddusi
, G. G. Cr
eame
r, N.
Ra
f
izadeh
, “M
achine
lea
r
nin
g
in
ene
rgy
economics
and
f
inance
: A r
e
v
ie
w,
Ener
g
y
Econo
mi
c
s
, v
ol
.
81
,
pp
.
709–727
,
2019
,
doi
:
h
tt
ps
:
//doi
.
o
rg
/10
.
1016/j
.
eneco
.
2019
.
05
.
006
.
[
99
]
F
.
Rundo
,
F
. Tr
en
t
a
, A. L.
di
S
t
allo
,
S
. B
a
tt
ia
t
o
, “M
achine
lea
r
nin
g
fo
r
quan
t
i
t
a
t
i
v
e
f
inance
applica
t
ions
: A
su
rv
e
y,A
pp
li
ed
Sc
i
ence
s
, v
ol
.
9
,
no
.
24
,
p
.
5574
,
2019
.
[
100
]
Z. L
in
, “M
odellin
g
and
fo
r
ecas
t
in
g
t
he
s
t
oc
k
ma
rk
e
t
v
ola
t
ili
t
y
of
SS
E
C
omposi
t
e
I
ndex
usin
g GA
R
CH
models
,
F
u
t
u
re
G
enerat
i
on
C
o
m
p
u
ter
S
y
s
te
m
s
, v
ol
.
79
,
pp
.
960–972
,
2018
.
[
101
]
M.
F
r
ömmel
, K. L
ampae
r
t
, “D
oes
f
r
equenc
y
ma
tt
e
r
fo
r
in
t
r
ada
y
t
echnical
t
r
adin
g
?
,
F
i
nance
Re
s
earc
h L
etter
s
, v
ol
.
18
,
pp
.
177–183
,
2016
.
[
102
]
M.
Sch
r
e
y
e
r, T.
Sa
tt
a
r
o
v, A. G
ie
r
bl
, B.
Reime
r, B
o
r
t
h
, “L
ea
r
nin
g
samplin
g
in
f
inancial
s
t
a
t
emen
t
audi
t
s
usin
g v
ec
t
o
r
quan
t
ised
v
a
r
ia
t
ional
au
t
oencode
r
neu
r
al
ne
t
w
o
rk
s
,
in
P
roceed
i
n
g
s
o
f
t
h
e
F
i
r
s
t
AC
M
I
nternat
i
ona
l C
on
f
erence
on
A
I
i
n
F
i
nance
,
2020
,
pp
.
1–
8
.
[
103
]
R
.
I
. W
ebb
, D. D.
R
y
u
, D. D.
R
y
u
,
J
. H
an
, “T
he
p
r
ice
impac
t
of
fu
t
u
r
es
t
r
ades
and
t
hei
r
in
t
r
ada
y
seasonali
t
y,
E
m
er
gi
n
g
Mar
k
et
s
Re
vi
e
w, v
ol
.
26
,
pp
.
80–
98
,
2016
,
doi
:
10
.
1016/j
.
emema
r.
2016
.
01
.
002
.
[
104
]
P
. B
acche
tt
a
, E. M
e
r
t
ens
, E. V
an
W
incoop
,
P
r
edic
t
abili
t
y
in
f
inancial
ma
rk
e
t
s
: W
ha
t
do
su
rv
e
y
expec
t
a
t
ions
t
ell
us?
,J
o
u
rna
l
o
f
I
nternat
i
ona
l
Mone
y
and
F
i
nance
, v
ol
.
28
,
no
.
3
,
pp
.
406–426
,
2009
.
[
105
]
L.
R
y
ll
,
S
.
Seidens
, “Ev
alua
t
in
g
t
he
pe
r
fo
r
mance
of
machine
lea
r
nin
g
al
g
o
r
i
t
hms
in
f
inancial
ma
rk
e
t
fo
r
ecas
t
in
g: A
comp
r
ehensi
v
e
su
rv
e
y,
ar
Xiv
prepr
i
nt
ar
Xiv:
1906.07786
,
2019
.
R
e
g
u
l
a
r
Iss
ue
-
1
4
7 -
[
106
]
A. M.
Ra
t
he
r, A. Ag
a
rw
al
, V.
Sas
t
ry,
Recu
rr
en
t
neu
r
al
ne
t
w
o
rk
and
a
h
y
b
r
id
model
fo
r
p
r
edic
t
ion
of
s
t
oc
k r
e
t
u
r
ns
,
E
x
pert
S
y
s
te
m
s
wi
t
h
A
pp
li
cat
i
on
s
, v
ol
.
42
,
no
.
6
,
pp
.
3234–3241
,
2015
.
[
107
]
J
.
P
.
Se
r
be
r
a
,
P
.
Pauma
r
d
, “T
he
fall
of
hi
g
h
-
f
r
equenc
y
t
r
adin
g: A
su
rv
e
y
of
compe
t
i
t
ion
and
p
r
o
f
i
t
s
,
Re
s
earc
h i
n
I
nternat
i
ona
l
B
u
s
i
ne
ss
and
F
i
nance
,
v
ol
.
36
,
pp
.
271–287
,
2016
,
doi
:
10
.
1016/j
.r
ibaf
.
2015
.
09
.
021
.
[
108
]
S
. A
lonso
-M
onsal
v
e
, A. L.
Suá
r
ez
-C
e
t
r
ulo
, A. C
e
rv
an
t
es
, D.
Quin
t
ana
,
“C
on
v
olu
t
ion
on
neu
r
al
ne
t
w
o
rk
s
fo
r
hi
g
h
-
f
r
equenc
y
t
r
end
p
r
edic
t
ion
of
c
ry
p
t
ocu
rr
enc
y
exchan
g
e
r
a
t
es
usin
g
t
echnical
indica
t
o
r
s
,
E
x
pert
S
y
s
te
m
s
wi
t
h A
pp
li
cat
i
on
s
,
1
2020
.
[
109
]
A. W. L
i
, G.
S
. B
as
t
os
, “
S
t
oc
k
ma
rk
e
t
fo
r
ecas
t
in
g
usin
g
deep
lea
r
nin
g
and
t
echnical
anal
y
sis
:
a
s
y
s
t
ema
t
ic
r
e
v
ie
w,
I
EEE
A
cce
ss
, v
ol
.
8
,
pp
.
185232–
185242
,
2020
.
[
110
]
J
.
Pa
t
el
,
S
.
Shah
,
P
. T
ha
kk
a
r, K. K
o
t
echa
,
P
r
edic
t
in
g
s
t
oc
k
ma
rk
e
t
index
usin
g
fusion
of
machine
lea
r
nin
g
t
echniques
,
E
x
pert
S
y
s
te
m
s
wi
t
h
A
pp
li
cat
i
on
s
, v
ol
.
42
,
pp
.
2162–2172
,
3
2015
.
[
111
]
M.
P
r
a
t
ama
,
S
. G. A
na
v
a
tt
i
,
P
.
P
. A
n
g
elo
v, E. L
u
g
hofe
r, “
P
AN
F
I
S
: A
no
v
el
inc
r
emen
t
al
lea
r
nin
g
machine
,
I
EEE
Tran
s
act
i
on
s
on
Ne
u
ra
l
Net
w
or
k
s
and
L
earn
i
n
g
S
y
s
te
m
s
, v
ol
.
25
,
pp
.
55–68
,
1
2014
.
[
112
]
G. W
idme
r, M. K
uba
t
, “L
ea
r
nin
g
in
t
he
p
r
esence
of
concep
t
d
r
if
t
and
hidden
con
t
ex
t
s
,
Mac
hi
ne
L
earn
i
n
g, v
ol
.
23
,
no
.
1
,
pp
.
69–101
,
1996
,
doi
:
10
.
1007/
B
F00116900
.
[
113
]
D.
Sahoo
,
Q
.
Pham
,
J
. L
u
,
S
. C. H
oi
,
Online
deep
lea
r
nin
g: L
ea
r
nin
g
deep
neu
r
al
ne
t
w
o
rk
s
on
t
he
f
l
y,
in
P
roceed
i
n
g
s
o
f
t
h
e
T
w
ent
y
-
Se
v
ent
h
I
nternat
i
ona
l J
o
i
nt
C
on
f
erence
on
A
rt
i
f
i
c
i
a
l
I
nte
llig
ence
I
JCA
I
2018
,
Jul
.
2018
,
pp
.
2660–2666
.
[
114
]
A. L. D.
Rossi
, B.
F
. D
e
Souza
, C.
Soa
r
es
, A.
de
L
eon
Fe
rr
ei
r
a
de
C
a
rv
alho
, C.
Ponce
, A g
uidance
of
da
t
a
s
t
r
eam
cha
r
ac
t
e
r
iza
t
ion
fo
r
me
t
a
-
lea
r
nin
g,
I
nte
llig
ent
D
ata
A
na
l
y
s
i
s
, v
ol
.
21
,
no
.
4
,
pp
.
1015–1035
,
2017
.
[
115
]
C. K
ädin
g, E.
Rodne
r, A.
F
r
e
y
t
a
g,
J
. D
enzle
r,
Fine
-
t
unin
g
deep
neu
r
al
ne
t
w
o
rk
s
in
con
t
inuous
lea
r
nin
g
scena
r
ios
,
in
C
o
m
p
u
ter
V
i
s
i
on
ACC
V
2016
W
or
k
s
h
op
s
, C
ham
,
2017
,
pp
.
588–605
,
Sp
r
in
g
e
r
I
n
t
e
r
na
t
ional
Publishin
g.
[
116
]
P
.
P
. A
n
g
elo
v,
E
m
p
i
r
i
ca
l A
pproac
h
to
Mac
hi
ne
L
earn
i
n
g.
Sp
r
in
g
e
r,
2017
.
[
117
]
B. Kr
a
w
cz
yk, L. L. M
in
k
u
,
J
. G
ama
,
J
.
S
t
efano
w
s
k
i
, M. W
oźnia
k, M.
W
ó
Z
nia
k, “E
nsemble
lea
r
nin
g
fo
r
da
t
a
s
t
r
eam
anal
y
sis
: A
su
rv
e
y,
I
n
f
or
m
at
i
on
F
u
s
i
on
, v
ol
.
37
,
pp
.
132–156
,
2017
.
[
118
]
R
. D. B
a
r
uah
,
P
. A
n
g
elo
v, “Ev
ol
v
in
g
fuzz
y
s
y
s
t
ems
fo
r
da
t
a
s
t
r
eams
: A
su
rv
e
y, Wil
e
y
I
nterd
i
s
c
i
p
li
nar
y
Re
vi
e
w
s
: D
ata
M
i
n
i
n
g
and
K
no
wl
ed
g
e
Di
s
co
v
er
y
, v
ol
.
1
,
pp
.
461–476
,
11
2011
.
[
119
]
M.
P
r
a
t
ama
,
S
. G. A
na
v
a
tt
i
, M.
Joo
, E. D. L
u
g
hofe
r,
pclass
: A
n
effec
t
i
v
e
classi
f
ie
r
fo
r
s
t
r
eamin
g
examples
,
I
EEE
Tran
s
act
i
on
s
on
F
uzz
y
S
y
s
te
m
s
,
v
ol
.
23
,
pp
.
369–386
, A
p
r.
2015
,
doi
:
10
.
1109/
T
F
UZZ.
2014
.
2312983
.
[
120
]
M. Á. A
bad
,
J
. B. G
omes
, E. M
enasal
v
as
, “
P
r
edic
t
in
g r
ecu
rr
in
g
concep
t
s
on
da
t
a
-
s
t
r
eams
b
y
means
of
a
me
t
a
-
model
and
a
fuzz
y
simila
r
i
t
y
func
t
ion
,
E
x
pert
S
y
s
te
m
s
Wi
t
h A
pp
li
cat
i
on
s
, v
ol
.
46
,
pp
.
87–105
,
2015
,
doi
:
10
.
1016/j
.
es
w
a
.
2015
.
10
.
022
.
[
121
]
J
.
Read
, “C
oncep
t
-
d
r
if
t
in
g
da
t
a
s
t
r
eams
a
r
e
t
ime
se
r
ies
;
t
he
case
fo
r
con
t
inuous
adap
t
a
t
ion
,
ar
Xiv
prepr
i
nt
ar
Xiv:
1810.02266
,
2018
.
[
122
]
H. M. G
omes
,
J
.
Read
, A. B
ife
t
,
J
.
P
. B
a
r
ddal
,
J
. G
ama
, “M
achine
lea
r
nin
g
fo
r
s
t
r
eamin
g
da
t
a
:
s
t
a
t
e
of
t
he
a
r
t
,
challen
g
es
,
and
oppo
r
t
uni
t
ies
,
AC
M
S
I
GKDD
E
x
p
l
orat
i
on
s
Ne
w
s
l
etter
, v
ol
.
21
,
no
.
2
,
pp
.
6–22
,
2019
,
doi
:
10
.
1145/3373464
.
3373470
.
[
123
]
X. Z
hen
g,
P
. L
i
, X. H
u
, K.
Yu
,
Semi
-
supe
rv
ised
classi
f
ica
t
ion
on
da
t
a
s
t
r
eams
w
i
t
h
r
ecu
rr
in
g
concep
t
d
r
if
t
and
concep
t
e
v
olu
t
ion
,K
no
wl
ed
g
e
-
Ba
s
ed
S
y
s
te
m
s
, v
ol
.
215
,
p
.
106749
,
2021
.
[
124
]
M.
P
r
a
t
ama
,
J
. L
u
, E. L
u
g
hofe
r, G. Z
han
g,
S
. A
na
v
a
tt
i
, “
Scaffoldin
g
t
y
pe
-
2
classi
f
ie
r
fo
r
inc
r
emen
t
al
lea
r
nin
g
unde
r
concep
t
d
r
if
t
s
,
Ne
u
roco
m
p
u
t
i
n
g,
v
ol
.
191
,
pp
.
304–329
,
2016
.
[
125
]
E. L
u
g
hofe
r,
P
. A
n
g
elo
v, “H
andlin
g
d
r
if
t
s
and
shif
t
s
in
on
-
line
da
t
a
s
t
r
eams
w
i
t
h
e
v
ol
v
in
g
fuzz
y
s
y
s
t
ems
, A
pp
l
.
So
f
t
C
o
m
p
u
t
.
, v
ol
.
11
,
pp
.
2057–2068
, M
a
r.
2011
,
doi
:
10
.
1016/j
.
asoc
.
2010
.
07
.
003
.
[
126
]
Ł. K
o
ry
c
k
i
, B. Kr
a
w
cz
yk, “
S
t
r
eamin
g
decision
t
r
ees
fo
r
lifelon
g
lea
r
nin
g,
in
J
o
i
nt
E
u
ropean
C
on
f
erence
on
Mac
hi
ne
L
earn
i
n
g
and
K
no
wl
ed
g
e
Di
s
co
v
er
y
i
n
D
ataba
s
e
s
,
2021
,
pp
.
502–518
,
Sp
r
in
g
e
r.
[
127
]
S
. U. D
in
,
J
.
Shao
, “E
xploi
t
in
g
e
v
ol
v
in
g
mic
r
o
-
clus
t
e
r
s
fo
r
da
t
a
s
t
r
eam
classi
f
ica
t
ion
w
i
t
h
eme
rg
in
g
class
de
t
ec
t
ion
,
I
n
f
or
m
at
i
on
Sc
i
ence
s
, v
ol
.
507
,
pp
.
404–420
,
2020
.
[
128
]
R
.
Szad
k
o
w
s
k
i
,
J
. Dr
chal
,
J
.
Fai
g
l
, “C
on
t
inuall
y
t
r
ained
life
-
lon
g
classi
f
ica
t
ion
,
Ne
u
ra
l C
o
m
p
u
t
i
n
g
and
A
pp
li
cat
i
on
s
,
pp
.
1–18
,
2021
.
w
v
[
129
]
N. K
asabo
v, D.
File
v, “Ev
ol
v
in
g
in
t
elli
g
en
t
s
y
s
t
ems
: M
e
t
hods
,
lea
r
nin
g,
applica
t
ions
,
in
2006
I
nternat
i
ona
l
S
y
m
po
s
ium
on
E
v
o
lvi
n
g
F
uzz
y
S
y
s
te
m
s
,
Sep
t
2006
,
pp
.
8–
18
.
[
130
]
S
.-
s
. Z
han
g,
J
.- . L
iu
, X. Z
uo
, “A
dap
t
i
v
e
online
inc
r
emen
t
al
lea
r
nin
g
fo
r
e
v
ol
v
in
g
da
t
a
s
t
r
eams
,A
pp
li
ed
So
f
t
C
o
m
p
u
t
i
n
g, v
ol
.
105
,
p
.
107255
,
2021
.
[
131
]
M. M
e
r
millod
, A. B
u
g
ais
k
a
,
P
. B
onin
, “T
he
s
t
abili
t
y-
plas
t
ici
t
y
dilemma
:
I
n
v
es
t
i
g
a
t
in
g
t
he
con
t
inuum
f
r
om
ca
t
as
t
r
ophic
fo
rg
e
tt
in
g
t
o
a
g
e
-
limi
t
ed
lea
r
nin
g
effec
t
s
,
F
ront
i
er
s
i
n
p
s
y
c
h
o
l
o
g
y
, v
ol
.
4
,
p
.
504
,
2013
.
[
132
]
E. L
u
g
hofe
r,
E
v
o
lvi
n
g fuzz
y
s
y
s
te
m
s
-m
et
h
odo
l
o
gi
e
s
,
ad
v
anced
concept
s
and
app
li
cat
i
on
s
, v
ol
.
53
.
Sp
r
in
g
e
r,
2011
.
[
133
]
M. C
a
r
nein
, H. Tr
au
t
mann
,
Op
t
imizin
g
da
t
a
s
t
r
eam
r
ep
r
esen
t
a
t
ion
: A
n
ex
t
ensi
v
e
su
rv
e
y
on
s
t
r
eam
clus
t
e
r
in
g
al
g
o
r
i
t
hms
,
B
u
s
i
ne
ss
&
I
n
f
or
m
at
i
on
S
y
s
te
m
s
En
gi
neer
i
n
g, v
ol
.
61
,
no
.
3
,
pp
.
277–297
,
2019
.
[
134
]
A. B
ife
t
,
R
. G
a
v
aldà
, G. H
olmes
, B.
Pfah
r
in
g
e
r,
Mac
hi
ne
l
earn
i
n
g f
or
data
s
trea
m
s
: wi
t
h
pract
i
ca
l
e
x
a
m
p
l
e
s
i
n
MO
A. M
I
T
p
r
ess
,
2018
.
[
135
]
K. N
ami
t
ha
, G.
San
t
hosh
K
uma
r, “L
ea
r
nin
g
in
t
he
p
r
esence
of
concep
t
r
ecu
rr
ence
in
da
t
a
s
t
r
eam
clus
t
e
r
in
g, J
o
u
rna
l
o
f
B
ig D
ata
, v
ol
.
7
,
p
.
75
,
D
ec
.
2020
,
doi
:
10
.
1186/s40537
-
020
-
00354
-
1
.
[
136
]
J
. G
ama
,
R
.
Sebas
t
iao
,
P
.
P
.
Rod
r
i
g
ues
,
On
e
v
alua
t
in
g
s
t
r
eam
lea
r
nin
g
al
g
o
r
i
t
hms
,
Mac
hi
ne
l
earn
i
n
g, v
ol
.
90
,
no
.
3
,
pp
.
317–346
,
2013
.
[
137
]
L.
Yan
g,
S
. M
c
C
lean
, M. D
onnell
y, K. B
u
rk
e
, K. K
han
, “D
e
t
ec
t
in
g
and
r
espondin
g
t
o
concep
t
d
r
if
t
in
business
p
r
ocesses
, Alg
or
i
t
hm
s
, v
ol
.
15
,
no
.
5
,
2022
,
doi
:
10
.
3390/a15050174
.
[
138
]
G.
Sa
t
eesh
B
abu
,
S
.
Su
r
esh
, G.-B. H
uan
g, “M
e
t
a
-
co
g
ni
t
i
v
e
N
eu
r
al
N
e
t
w
o
rk
fo
r
classi
f
ica
t
ion
p
r
oblems
in
a
sequen
t
ial
lea
r
nin
g
f
r
ame
w
o
rk,
Ne
u
roco
m
p
u
t
i
n
g, v
ol
.
81
,
pp
.
86–96
,
2011
.
[
139
]
A. M
aslo
, M.
Pecheniz
k
i
y,
I
. Ž
liobai
t
e
˙
, T. K
ä
rkk
äinen
, “M
odellin
g
r
ecu
rr
en
t
e
v
en
t
s
fo
r
imp
r
o
v
in
g
online
chan
g
e
de
t
ec
t
ion
,
in
P
roceed
i
n
g
s
o
f
t
h
e
2016
S
I
A
M
I
nternat
i
ona
l C
on
f
erence
on
D
ata
M
i
n
i
n
g,
2016
,
pp
.
549–
557
,
S
I
AM.
[
140
]
P
.
P
. A
n
g
elo
v, D.
P
.
File
v, A
n
app
r
oach
t
o
online
iden
t
i
f
ica
t
ion
of
t
a
k
a
g
i
-
su
g
eno
fuzz
y
models
,
Tran
s.
S
y
s.
Man
C
y
ber
.
P
art
B
, v
ol
.
34
,
pp
.
484–498
,
Feb
.
2004
,
doi
:
10
.
1109/
T
S
MCB.
2003
.
817053
.
[
141
]
C. W. C
hiu
, L. L. M
in
k
u
, A
di
v
e
r
si
t
y
f
r
ame
w
o
rk
fo
r
dealin
g w
i
t
h
mul
t
iple
t
y
pes
of
concep
t
d
r
if
t
based
on
clus
t
e
r
in
g
in
t
he
model
space
,
I
EEE
Tran
s
act
i
on
s
on
Ne
u
ra
l
Net
w
or
k
s
and
L
earn
i
n
g
S
y
s
te
m
s
,
2020
.
[
142
]
A. L.
Suá
r
ez
-C
e
t
r
ulo
, A. C
e
rv
an
t
es
, A
n
online
classi
f
ica
t
ion
al
g
o
r
i
t
hm
fo
r
la
rg
e
scale
da
t
a
s
t
r
eams
:
i
g
n
g
s
v
m
,
Ne
u
roco
m
p
u
t
i
n
g, v
ol
.
262
,
pp
.
67–
76
,
2017
,
doi
:
h
tt
ps
:
//doi
.
o
rg
/10
.
1016/j
.
neucom
.
2016
.
12
.
093
.
[
143
]
E. L
u
g
hofe
r, C. C
e
r
nuda
,
S
. K
inde
r
mann
, M.
P
r
a
t
ama
, “G
ene
r
alized
sma
r
t
e
v
ol
v
in
g
fuzz
y
s
y
s
t
ems
,
E
v
o
lvi
n
g
S
y
s
te
m
s
, v
ol
.
6
,
no
.
4
,
pp
.
269–292
,
2015
,
doi
:
10
.
1007/s12530
-
015
-
9132
-
6
.
[
144
]
P
. E. T
sinaslanidis
, D. K
u
g
ium
t
zis
, A
p
r
edic
t
ion
scheme
usin
g
pe
r
cep
t
uall
y
impo
r
t
an
t
poin
t
s
and
d
y
namic
t
ime
w
a
r
pin
g,
E
x
pert
S
y
s
te
m
s
wi
t
h A
pp
li
cat
i
on
s
, v
ol
.
41
,
no
.
15
,
pp
.
6848–6860
,
2014
,
doi
:
h
tt
ps
:
//doi
.
o
rg
/10
.
1016/j
.
es
w
a
.
2014
.
04
.
028
.
[
145
]
C. M
a
r
t
in
, D.
Quin
t
ana
,
P
.
I
sasi
, “Dy
namic
g
ene
r
a
t
ion
of
in
v
es
t
men
t
r
ecommenda
t
ions
usin
g gr
amma
t
ical
e
v
olu
t
ion
.,
I
nternat
i
ona
l J
o
u
rna
l
o
f
I
nteract
iv
e
M
ul
t
im
ed
i
a
&
A
rt
i
f
i
c
i
a
l
I
nte
llig
ence
, v
ol
.
6
,
no
.
6
,
2021
.
[
146
]
B. Gr
ün
, “M
odel
-B
ased
C
lus
t
e
r
in
g,
Handboo
k
o
f
M
ix
t
u
re
A
na
l
y
s
i
s
,
pp
.
157–192
,
Feb
.
2019
,
doi
:
10
.
1201/9780429055911
-
8
.
[
147
]
P
. D. M
c
N
icholas
, “M
odel
-
based
clus
t
e
r
in
g,J
o
u
rna
l
o
f Cl
a
ss
i
f
i
cat
i
on
, v
ol
.
33
,
no
.
3
,
pp
.
331–373
,
2016
.
[
148
]
F
. D
ellae
r
t
, “T
he
expec
t
a
t
ion
maximiza
t
ion
al
g
o
r
i
t
hm
,” G
eo
rg
ia
I
ns
t
i
t
u
t
e
of
T
echnolo
gy,
2002
.
[
149
]
J
. V
anscho
r
en
,
J
. N. V
an
Rijn
, B. B
ischl
, L. T
o
rg
o
,
Openml
:
ne
t
w
o
rk
ed
science
in
machine
lea
r
nin
g,AC
M
S
I
GKDD
E
x
p
l
orat
i
on
s
Ne
w
s
l
etter
, v
ol
.
15
,
no
.
2
,
pp
.
49–60
,
2014
.
[
150
]
P
.
P
. A
n
g
elo
v, X. Z
hou
, “Ev
ol
v
in
g
fuzz
y-r
ule
-
based
classi
f
ie
r
s
f
r
om
da
t
a
s
t
r
eams
,
I
EEE
Tran
s
act
i
on
s
on
F
uzz
y
S
y
s
te
m
s
, v
ol
.
16
,
pp
.
1462–1475
, D
ec
.
2008
,
doi
:
10
.
1109/
T
F
UZZ.
2008
.
925904
.
[
151
]
O
. C
o
r
dón
,
F
. G
omide
,
F
. H
e
rr
e
r
a
,
F
. H
offmann
, L. M
a
g
dalena
, “T
en
y
ea
r
s
of
g
ene
t
ic
fuzz
y
s
y
s
t
ems
: C
u
rr
en
t
f
r
ame
w
o
rk
and
ne
w
t
r
ends
,
in
F
uzz
y
Set
s
and
S
y
s
te
m
s
, v
ol
.
141
,
2004
,
pp
.
5–31
.
I
ntern
a
t
i
on
a
l
J
ourn
a
l
of
I
nter
ac
t
i
v
e
M
u
l
t
i
med
i
a a
nd
Art
i
f
i
c
i
a
l
I
nte
lli
g
en
c
e,
Vo
l
.
9
,
N
º
1
-
1
48
-
Andr
é
s
L
.
S
uár
e
z
-C
et
ru
l
o
Andrés received his BSc and MSc in Computer Science at
Carlos III of Madrid (Spain) in 2013 and 2014, respectively.
He received his PhD in Computer Science at University
Carlos III of Madrid in 2022. He is currently a Data Science
Architect at Ireland’s Nationa Centre for Applied AI, based
online machine learning for data streams, regime changes
at University College Dublin.
l
His current interests focus on
in
f
inancial markets, deep learning, transformers and generative models.
Dav
i
d
Q
u
i
n
t
ana
Admin stration and Computer Science. He ha an M.S.
in Intelligent Systems from Universidad Carlos III de
Madrid and a PhD in F nance from Universidad Ponti
f
icia
the Computer Science Department at University Carlos III
of Madrid. There, he is part of the bio-inspired algorit ms
group EVANNAI. His current research interests are mainly fo
David
i
Quintana holds Bachelor’s degrees in
s
Business
Comillas (ICADE). He
i
is currently Associate Professor at
applications of Computational Intelligence in
f
inance and economics.
cused
h
on
A
l
e
j
andr
o
C
e
rvan
te
s
Graduated as Telecommunications Engineer at Universidad
Politecnica of Madrid (Spain), in 1993. He received his
PhD in Computer Science at University Carlos III of
Superior de Ingeniería y Tecnología in UNIR (Universidad
algorithms for classi
f
ication of non-stationary data, large
multi-objective optimization problems, swarm in elligence algorithms and deep
Madrid in 2007. He is currently a professor at the Escuela
Internacional de la Rioja). His interests include bio-inspired
machine learning for meteorological forecasting,
t
aeronautics and astrophysics.